Ab initio quantum molecular dynamics method based on the … · 2020. 4. 2. · AB fivmo QUANTUM...
Transcript of Ab initio quantum molecular dynamics method based on the … · 2020. 4. 2. · AB fivmo QUANTUM...
Ab initio quantum molecular dynamics methodbased on the restricted-path integral: Application
to electron plasma and an alkali metal
Item Type text; Dissertation-Reproduction (electronic)
Authors Oh, Ki-Dong
Publisher The University of Arizona.
Rights Copyright © is held by the author. Digital access to this materialis made possible by the University Libraries, University of Arizona.Further transmission, reproduction or presentation (such aspublic display or performance) of protected items is prohibitedexcept with permission of the author.
Download date 24/02/2021 08:54:29
Link to Item http://hdl.handle.net/10150/284195
INFORMATION TO USERS
This manuscript has been reproduced from the microfilm master. UMI
films the text directly fix)m the original or copy submitted. Thus, some
thesis and dissertation copies are in typewriter &ce, while others may be
fi-om any type of computer printer.
The quality of this reproduction is dependent upon the quality of the
copy submitted. Broken or indistinct print, colored or poor quality
illustrations and photographs, print bleedthrough, substandard margins,
and improper aligmnent can adversely affect reproduction.
In the unlikely event that the author did not send UMI a complete
manuscript and there are missing pages, these will be noted. Also, if
unauthorized copyright material had to be removed, a note will indicate
the deletion.
Oversize materials (e.g., maps, drawings, charts) are reproduced by
sectioning the original, banning at the upper left-hand comer and
continuing fi'om left to right in equal sections with small overlaps. Each
original is also photognq)hed in one exposure and is included in reduced
form at the back of the book.
Photographs included in the original manuscript have been reproduced
xerographically in this copy. Higher quality 6" x 9" black and white
photogr^hic prints are available for any photographs or illustrations
appearing in this copy for an additional charge. Contact UMI directly to
order.
UMI A Bdl & Howell Infimnation Conqai
300 Noith Zeeb Road, Ann Aibor MI 48106>1346 USA 313/761-4700 800/521-0600
AB fivmo QUANTUM MOLECULAR DYNAMICS METHOD BASED ON THE
RESTRICTED PATH INTEGRAL: APPLICATION TO ELECTRON PLASMA AND
AN ALKALI METAL
by
Ki-Dong Oh
A Dissertation Submitted to the Faculty of the
DEPARTMENT OF PHYSICS
In Partial Fulfillment of the Requirements
For the Degree of
Doctor of Philosophy
In the Graduate College
THE UNIVERSITY OF ARIZONA
19 9 9
DMI Niunber: 9927497
UMI Microform 9911 ASH Copyright 1999, by UMI Company. All rights reserved.
This microform edition is protected against unauthorized copying under Title 17, United States Code.
UMI 300 North Zeeb Road Ann Arbor, MI 48103
2
THE UNIVERSITY OF ARIZONA ® GRADUATE COLLEGE
As members of the Final Examination Committee, we certify that we have
read the dissertation prepared by KI-DONG OH
entitled AR TNITTO miANTtlM MOLECULAR DYNAMICS METHOD BASED ON THE
RF.STRTCTKD PATH TNTKfiRAL: APPLICATION TO ELECTRON PLASMA
AND AN ALKALI METAL.
and recommend that it be accepted as fulfilling the dissertation
requirement for the Degree of Doctor of Philosophy
Robert H.,Parmenter Date nprember 23. 1998
December 23» 1998 loval W. Stark Date
December 23» 1998 Robert H. Chambers Date
^ C0)M^9J^iv\ i-L f i - j / Laurence Mclntyre Date
December 23» 1998 e A. Deymler Date
Final approval and acceptance of this dissertation is contingent upon the candidate's submission of the final copy of the dissertation to the Graduate College.
I hereby certify that I have read this dissertation prepared under my direction and recommend that it be accepted as fulfilling the dissertation requirement.
- DPfPinher 23. 1998 Dlssertatl(5n Director Pierre A. Deymler Date
December 23, 1998 -Dissertation Director Robert H. Chambers Date
3
STATEMENT BY AUTHOR
This dissertation has been submitted in partial fulfillment of requirements for an advanced degree at The University of Arizona and is deposited in the University Library to be made available to borrowers under rules of the library.
Brief quotations from this dissertation are allowable without special permission, provided that accurate acknowledgment of source is made. Requests for permission for extended quotation from or reproduction of this manuscript in whole or in part may be granted by the head of the major department or the Dean of the Graduate College when in his or her judgment the proposed use of the material is in the interests of scholarship. In all other instances, however, permission must be obtained from the author.
SIGNED
4
ACKNOWLEDGEMENTS
It is a great pleasure to thank to my advisor Prof. Pierre Deymier for his
considerate guidance and help from the beginning to the end of this work. Without his
encouragement amd patience, this work whould not have been possible. He shared his
expert knowledge and experiences, and made vital suggestions whenever I needed advice
and encouragement, amd I am extremely indebted to him.
A special thanks to my friend Frank Cheme with whom I shared the ssune lab,
same frustrations, and same happiness. The computer facilities in the Telecommunica
tion Center at the Cornell University and the CCIT, Physics Department, and Material
Science and Enginieering Department at the University of Arizona were essential for this
work. I am greatfiil to Dr. Jadme Combariza for granting me generous computing time
and helpful tips of MPI programs.
Finally, I thauak my family, especially my wife Kim Soojung who gave me un
conditional support and care over the years. Most of all I dedicate this dissertation to
my father and mother who are praying for us everyday.
5
TABLE OF CONTENTS
LIST OF TABLES 8
LIST OF FIGURES 9
ABSTRACT 13
1 INTRODUCTION 15
2 THEORY 18
2.1 Feynman Path Integral 18
2.1.1 Partition function for a single particle 18
2.1.2 Systems of interacting particles obeying Maxwell-Boltzmann statis
tics 23
2.1.3 Two-Electron system 25
2.1.4 Many-Electron system 29
2.2 Path Integral with Non-local Exchange Using the Mean Field Approximation 33
2.3 Restricted Path Integral Method 35
2.4 Classical Isomorphism for Many-body fermionic system 38
2.5 Molecular Dynamics 41
6
2.5.1 Background 41
2.5.2 Restricted Path Integral Molecular Dynamics Method 46
2.6 Physical Quantities 48
2.6.1 Average of Physical Quantities 48
2.6.2 Energy Estimator 48
2.6.3 Correlation Functions 50
3 PRACTICAL IMPLEMENTATION 52
3.1 Treatment of Classical Potential Energy 52
3.1.1 Electron Plasma System 52
3.1.2 Alkali Metal System 55
3.2 .Algorithm of Molecular Dynamics 57
3.2.1 Overview 57
3.2.2 Creation of the Initial Configurations 60
3.3 Periodic Boundary Conditions in Path Integral MD 62
3.4 Calculation of Quantum Effects 66
3.4.1 Evaluation of det(£'^''''^) of the Effective Exchange potential with
PBC 66
3.4.2 Effective Force Calculation 70
3.5 Parallel Computation 71
4 APPLICATIONS AND RESULTS 76
I
4.1 Electron Plasma 76
4.1.1 Model System 76
4.1.2 Results 79
4.2 Solid/Liquid Transition of Alkali Metal 87
4.2.1 Model System 87
4.2.2 Results and Discussion 89
5 CONCLUSIONS AND FUTURE WORK 109
6 APPENDICES 111
6.1 .A.PPENDIX A : FREE R\RTICLE PROPAGATOR Ill
6.2 .\PPENDIX B : EXCHANGE KINETIC ENERGY ESTIMATOR ... 112
6.3 APPENDIX C : EXCHANGE FORCE CALCULATION 113
REFERENCES 115
8
LIST OF TABLES
4.1 Kinetic (A) and potential (B) energies per electron for various electron
densities. Kinetic energies per electron of current model are the average
values. Potential energies per electron are linearly fitted by using PE = a
-I- bT 83
4.2 Specific heat of the solid and the liquid, and latent heat. The experimental
values are at the data book(CRC). (1) and (2) in (A) are measure at T =
lOOK and 298K 90
9
LIST OF FIGURES
2.1 Necklace representation of a single particle for P = A. The number labels
indicate the different imaginary times 21
2.2 Exchange process between two electrons 27
2.3 Three-cycle exchange process among three identicle particles. Solid lines
and dashed lines represent two different exchange processes 31
2.4 Local(a) and non-local(b) exchange processes between two electrons. The
number labels indicate the imaginary times 35
2.5 Phase diagram of the sign of determinant. The diagonal line is x • y =
0. All unfilled dots correspond to positive density matrixes and the filled
dot has a density matrix with a negative value 39
2.6 Basic simulation cell(thick solid line) and its 8 image cells in two dimen
sional system with PBS. The size of the basic cell is L. The radius of
the dotted circle is L/2. The dotted circle area is the effective range of
interaction 45
3.7 Coulomb interaction between two electrons i and j at different imainary
times 53
3.8 Empty core pseudopotential model for the ion-electron interaction. The
sum of the potentials (b), (c), and (d) is equal to (a) 57
10
3.9 Initial necklace configurations for two electrons in an electron plasma at
T = I300K and r, = .5. Only 1/3 of the beaxls {P = 480) denoted by filled
circles are shown in xy-plain 62
3.10 Beads of an electron necklace in the simulation cell and the image ceils.
Beads (1 and 4) are in the simulation cell and beads (2 and 3) which belong
to the same necklace are in the image cell. In the usual MD method, we
use the filled circles in (a) to calculate distances between beads. In (b),
the necklacp is reronstructed after translating beads (2 and 3) by L from
the left 63
3.11 PBC diagram with two electrons. The smaller filled dots represent the /th
beads of the electron j and the larger circles represent the berds of of the
electron i. We assume that the beads denoted by the large filled circles
are in the simulation cell 65
3.12 Algorithm for the calculation of the determinant det(£'^''-'^) 69
3.13 Sequantial job order in a parallel computer 73
3.14 Scalability of the parallel caiculation. CPU time scale is normalized to 1
for the serial calculation with one processor. .\ll time is measured at the
IBM SP2 in the Telecommunication Center at Cornell University 74
4.15 V'®"'' and potential energy in reduced units as functions of time steps. In
both cases, the thick lines and the thin dotted lines refer to skip = 10 and
skip = 0, respectively. 78
4.16 Kinetic energy of electron plasmas as function of number of beads in the
necklace representation of quantum particles. The circles and squares
refer to the high density (r, = 5, T=1800K) and medium density (r, =
7.5, r=700K) 80
11
4.17 Potential energy as functions of number of beads. The circles and squares
refer to the high density (r, = 5, r=1800K) and medium density (r, =
7.5, r=700K) 81
4.18 Kinetic energy as functions of temperature. The electron plasma with
ra=o, ra=7.5 and r5=10 are refered to by circles, squares and triangles,
respectively. The horizontal thick dashed lines correspond to the energies
of Ceperley[12,13]The thin dotted lines indicate the Hartree-Fock energies. 84
4.19 Potential energy as functions of temperature. The electron plasma with
ra=5, r,=7.5 and rj=10 are refered to by circles, squares and triangles,
respectively. The horizontal thick dashed lines correspond to the energies
of Ceperley[12,13]. The thin dotted lines indicate the Hartree-Fock en
ergies. For both types of lines, r^^o, ra=7.5 and rj=10 are represented
from the top to the bottom 85
4.20 Iso-spin and hetero-spin electron-electron pair distributions for the high
density (r, = 5) electron plasma at T=1300 K (solid lines), T=1800 K
(dotted lines) and T=2300 K (dashed lines) 86
4.21 Running averages of electron kinetic energies at T = 273K and T = lOK.
The standard deviations are 0.003 (eV/electron) and 0.005 (eV/electron)
at T = 273K and at T = lOK, respectively. 97
4.22 Convergence of electron kinetic energy with respect to the number of beads
(P) in potassium. r,o„ = 273 K and Teu = 1300 K. Nion = ^eie = 54. . 98
4.23 Total energy of the potassium model versus temperature. The lines are
fits to the data in the low and high temperature regions 99
4.24 Various contributions to the total energy of the potassium system as func
tions of temperature 100
12
4.25 Ion pair distribution functions at the different temperatures of 10K(thick
solid line), 76K(thick dotted line), 150K(dashed line), 248K(thin solid
line), 273K(thin dotted line), and 298K(thick long dashed line) 101
4.26 Trajectories of the potassium ions at (a) T=10K, (b) T=76K, (c) T=248K.
and (d) T=298K 102
4.27 Mean square displacement(MSD) of potassium ions as a function of time
and temperature 103
4.28 Vibrational amplitude of the potassium ions as function of the ion tem
perature. The dotted line is a linear fit to the data except for the last
point 104
4.29 Normalized velocity autocorrelation function(NVAF) and associated power
spectrum for crystalline potassium(T=10K) and liquid metal(T=273K
and T=298K). The insert in the T=10K power spectrum is the exper
imentally deduced phonon density of state at T=9K of a reference [19]. - 105
4.30 A 2D projection of the electron necklace (open circle) with potassium ions
(larsje filled circle) at T = 298K and T = lOK. The frame represents the
simulation cell 106
4.31 Partial electron-electron pair correlation functions. The solid lines and
dotted lines refer to the crystal at T=10K and the liquid at T=273K,
respectively 107
4.32 Ion-electron pair distribution functions at several temperatures 108
13
ABSTRACT
We develop a new Quantum Molecular Dynamics simulation method. The method is
based on the discretized path integral representaion of quantum mechanics. In this rep
resentation, a quantum particle is isomorphic to a closed polymer chain. The problem of
the indistinguishability between quantum particles is tackled with a non-local exchange
potential. When the exact density matrix of the quantum particles is used, the exchange
potential is exact. However we use a high temperature approximation to the density
matrix and the exchange potential is only approximate. This new quantum molecular
dynamics method allows the simulation of collections of quantum particles at finite tem
perature. Our algorithm can be made to scaJe linearly with the number of quantum
states on which the density matrix is projected. Therefore, it can be optimized to run
efficiently on parallel computers.
We apply this method to the simulation of the electron plasma in 3-dimensions
with different densities (r, = 5.0, 7.5, and lO.O) at various temperatures. Under these
conditions, the electron plasma are at the border of the degenerate and the semi-
degenerate regimes. The kinetic and potential energies are calculated and compared
with results for similar systems simulated with a variational Monte Carlo method. Both
results show good agreements with each other at aJl the densities studied.
The quantum path integral molecular dynamics is also employed to study the
effect of temperature on the electronic and atomic structural properties of liquid and crys
talline alkali metai, namely potassium. In these simulations, ions and valence electrons
are treated as classical and quantum particles, respectively. The simple metal undergoes
a phase transformation upon heating. Calculated dynamic properties indicate that the
14
atomic motion changes from a vibrational to a diffusive character identifying the trans
formation as melting. Calculated structural properties further confirm the nature of the
transformation. Ionic vibrations in the crystal state and the loss of long range order
during melting modify the electronic structure and in particular localize the electrons
inside and at the border of the ion core.
15
CHAPTER 1
rNTRODUCTION
Modeling and simulation have become a vital part of materials research. Modeling and
simulation techniques are maturing to the point where they offer hope for a practical and
reliable approeich for the study of real materials. The development of materials models
has evolved from the infancy of specific empirical descriptions, to highly accurate and
sophisticated representations based on first principle calculations. In the field of Ab-
initio molecular dynamics method, the method of Car and Parrinello[ll, 74], based
on the Density Functional Theory(DFT) has enjoyed a great popularity over recent
years. DFT molecular dynamics has been employed to investigate a very large number
of problems from condensed matter to chemistry to biology [65]. In constrast, applications
of molecular dynamics simulations using the discretized path-integral[23] representation
of quantum particles have been limited mostly to the simulation of systems containing
a small number of quantum degree of freedom (such as in the solvation of a single
quantum particle in a classical fluid[64]) or to problems where quantum exchange is
not dominant[50]. We should aiso mention the path-integral based method of Alavi and
Frenkel that allows for the calculation of the grand canonical partition function of fermion
systems[l]. With this method the fermion sign problem in the evaluation of the partition
function is solved exactly in the case of non-interacting fermions. When combined with
DFT, this method provides a means of doing ab initio molecular dynamics of systems
with interacting high temperature electrons[2].
Progresses in the simulation of fermionic systems by path-integral Monte Carlo
16
method[34, 35, 36, 15, 17, 95] have opened the way toward the implementation of a path-
integral based finite temperature ab initio molecular dynamics method(PIMD). In this
study, we describe such a molecular dynamics method applicable to the simulation of
many-fermion systems at finite temperatures. The method is based on (a) the discretized
path integral representation of quantum particles as closed polymeric chains of classical
particles (or beads) coupled through harmonic springs[23], (b) the treatment of quantum
exchange as crosslinking of the chains[18], (c) the non-locality of crosslinking (exchange)
along the imaginary time chains[34,35, 36], and (d) the restricted path integral[14, 15. 16]
to resolve the problem of negative weights to the partition function resulting from the
crosslinidng of even numbers of quantum particles.
The present PFMD is initially applied to the description of one-component
plasma, which consists of the electron gas with a uniform neutralizing background, at
the border of the degenerate and semi-degenerate regimes where the ratio of the tem
perature to the Fermi temperature(7V) « 0.1. The electron plasma is the first focus of
our investigation because it is the simplest many-body fermionic system. It has been
extensively studied via path-integral, variational, and diffusion Monte Carlo methods
since the calculation of the equation of states of a Fermi one-component plasma such
as the interacting electron gas is a problem of fundamental practical importance as one
uses its properties in the density functional theory. The one-component plasma is also
a good prototype system as there exists a large amount of theoretical and numerical
data on its equation of state. The zero-temperature perterbative expansion of the en
ergy of a three-dimensional uniform electron plasma in the high density limit, where
''j 1 (r, = r/oQ where r is the electron sphere radius and qq is the Bohr radius), was
calculated theoretically quite some time ago[28]. Accurate variational Monte Carlo cal
culations have extended the zero-temperature equation of states of the degenerate Fermi
one-component plasma to a wide range of lower densities from r, = 1 to 500[12, 13]. The
exchange-correlation free energy has been subsequently calculated to encompass the full
range of thermal degeneracy[20, 66, 79].
17
After showing that the PEMD contains the necessary ingredients to simulate elec
tron plasma up to metal densities at finite temperatures, we apply the PIMD to a simple
alkali metal, namely potassium K. We chose potassium because (1) it is a prototype
free-electron metal, (2) there exist experimental data for the pair correlation function of
a liquid potassium[89] and the power spectrum of crystalline potassium at 9A'[19], (3)
DFT molecular dynamics has had problems with metals when electrons leave the Born-
Oppenheimer surface and therefore violate one of the basic assumption of the method.
This problem has been solved technically in an ad-hoc manner with the introduction of
appropriate thermostats for the electronic and ionic degrees of freedom[8]. In the simple
metal case, the discretized restricted path integral representation of electron is the same
as that of the electron plasma. Classical ionic degrees of freedom representing potassium
ions are added to the model. We firstly show that the PIMD successfully models the
body centered crystal structure of the solid state of potassium at low temperature. Upon
increasing the temperature, the solid state transforms to a liquid. The predicted melting
temperature is below the experimental value. This deviation is assigned to a short-range
approximate form used in place of the usual long-range Coulomb potential in order to re
duce computationcil time. The phase transformation is characterized thermodynamically
via energies, structually via pair distribution functions as well as dynamically via the
mean square displacements and the vibrational power spectra. Vibrations in the crystal
appear to induce some localization in the electron density. .A.s melting takes place, the
electronic structure responds to the loss of long range order in the atomic structure by
additionaJ localization.
We introduce the development of the path-integral molecular dynamics from a
one-body system to a non-localized many-fermion system in Chapter 2. In Chapter 3, we
explain the method of PIMD in further details and describe its practical implementation.
In Chapter 4, we present the results of PIMD calculation of electron plasma and an alkali
metal. Finally, in Chapter 5, we draw conclusions concerning the applicability of the
PIMD method to some other materials systems and suggest some future work.
18
CHAPTER 2
THEORY
2.1 Feynman Path Integral
2.1.1 Peirtition function for a single particle
Since Feynman[22] introduced the path integral of a quantum system, it has been well
developed[24, 25, 67] and applied to many-body systems[15, 18, 39, 58, 59, 60, 64]. The
basic idea of the Feynman path integral is to break a finite time interval into infinitesimal
time steps and then evaluate the matrix element of the propagation operator for each
step. In quantum statistical mechanics, all static properties and dynamic properties
of a system in thermal equilibrium are specified from the thermal density matrix. If
we work in the canonical ensemble, which is a system of fixed number of particles in a
fixed volume in equilibrium with a thermal reservoir, the probability of observing a state
with energy E is proportional to where ks is Boltzmann's constant and T is
the temperature. Let us consider a single particle system governed by the Hamiltonian
operator H. The partition function of this system may be written
Z = Tre ' ^"
= J dr <r|e~''^|r> (2.1)
where /3 is l / k sT and |r> is the exact eigenstate of H. In the path integral formalism for
many-body systems, we normally represent the matrix element in the partition function
as a density matrix, p{r{,rj;/3). The density matrix is defined as
p(r,-,rj;/3) = <r,|e- |rj> . (2.2)
19
Then the partition function can be rewritten as the trace of the density matrix
Z = J drp{r , r ; l3 ) . (2.3)
Before proceeding with the development of approximate forms for the density
matrix, we will first consider the matrix element in real time, f, for physicai clarity. By
substituting i{tf — ti)/h for 13, the density matrix p{ri,rf;0) for a particle governed by
the Hamiltonian, H, becomes
K(r j , t f ; r i , t i ) = <r (2.4)
where and t j are an initial time and a final time, respectively. The matrix element in
real time, K(r j,tf-ri,ti), which is the so called Kernel [23], is a solution of a real time
dependent Schrodinger Equation.
wherein the Hamiltonian H/ operates on the variables r/ and tj only. By analogy, the
density matrix in imaginary time is a solution of an equation of the form
dp -T3 = '2.6)
Eq.( 2.6) is a diffusion like equation. This fact will become important when we introduce
the restricted path integral and in particular, when we consider boundary conditions on
the density matrix. The kernel K{rj,tf\ri,ti) obeys the superposition principle, since
it is an exact solution of the Schrodinger equation in real time. By the superposition
principle, we mean that
= j drK{r j , t f , r , t )K{r , t ; r i , t i ) (2.7)
at any time t , where t i < t < t f . Eq. ( 2.7) indicates that one may calculate the matrix
element to any desired degree of accuracy for infinitesimal time interval, although the
matrix element, or the kernel A'(r/,i/;r,-,f,), cannot be calculated exactly for a finite
time interval, tj — ti. This is the basic idea of the Feynman path integral. In other words,
20
we calculate a matrix element only for each infinitesimal time interval after breaking a
finite time into infinitesimal intervals. The value for the finite time interval can be obtain
from the results of the evaluations for ail infinitesimal time intervals. Similarly, we may
evaluate the thermal density matrix p(r{, rj;/3) with appropriate accuracy if we divide
a finite temperature term 0 in P infinitesimal intervals, where P —¥ oo.
To evaluate the thermal density matrix, we will consider a set of P
different configurationa l s t a t e s , { | r , - > ; i = 1 , P} , where each s t a t e i s an e igens ta t e o f H.
With the relation
or
^ -0H ^
where £ —QjP, the partition function becomes
Z = J dr <rle-'^e-'^ • • • e-'"\r> .
Using the closure relation of the eigenstates |r,> of H,
J dvi Ir.xril = 1
and projecting the particle on (P— 1) intermediate states, the partition function can be
written as
^ = Jdridr2...drp <ri|e"''^|r2><r2|e~''^|r3> • • • <rp|e"'^|rp+i>, (2.8)
or
Z = j dr idr2 . . . d rpp{r i , r2 ;e )p{r2 , r3 ;e ) • •p{rp , rp^ . i ;€ ) (2.9)
where ri = rp^i indicates complete closure. Figure (2.1) illustrates a complete necklace
of a single particle for P = 4.
Each matrix element or each density matrix in the above relation represents
a propagator from one state to another state for infinitesimal imaginary time or very
21
«(.P)
Figure 2.1: Necklace representation of a single particle for P = 4. The number labels indicate the different imaginary times.
small deviation of temperature. In other words, the density matrix of a single particle
is connected to look like a polymeric necklace consisting of P beads. To evaluate the
density matrix of a single particle for infinitesimal imaginary time, we assume PT is
very large. Then we can adopt Trotter's second order appoximation[3]. According to
the Trotter formula[3-5],we have
e-"" = = Urn (2.10) P-+-00
and
g-e ( t+V ' ) smal l € (2.11)
where the Hamiltonian f l = T ->rV. T and V are a kinetic energy operator and a potential
energy operator , respectively. With Eqn.( 2.11), we can approximate the exact density
matrix by the product of the density matrix for T and the density matrix for V. The
error of this appoximation is of order of
Now we are going to evaluate the density matrix for infinitesimal imag
inary time step c = l3/P. Let us assume that the Hamitonian H = T + K,where
T = —p^ jlm and V = K(f) is a local potential energy operator. Introducing a complete
set of momentum states, |p„> and using Eqn. ( 2.11), the density matrix becomes
A>(rn,'*„+i;6) = <r„|e~'=^|r„+i>
22
= J dPn<rn\Pn><PnW '^\rn+l> = yrfp„<r„|p„><pJe-^'/2-e-'^|r„+i>+C7(ee). (2.12)
After performing the momentun operator to |p„> and the local potential operator to
lrn+i>, then we obtain
/>(r„,rn+i;e) « (2.13)
where we define Po(»'n, ^n+i; c) as the density matrix of a free particle (also called the
free particle propagator), as
Po(r„,r„+i;e) = <r„|e"'=^|r„+i> (2.14)
or
Po(rn,r„+i;€) = J dp„ <rnlP„><Pnkn+I> e ^2.15)
Using the Gaussian integral,the free particle propagator becomes
PoK,r„«;^) = (^) ' exp . (2.16)
We give more details of the derivation of Eq. ( 2.16) in the Appendix A. From Eq. (2.9)
with £ = /?/P, the partition function of a single particle can be written as
r P , g Z = Yl 'drnp{r ,„r„+i ;—)
n=l ^
~ (2.17) n=l ^
or
with
(2.18) 11=1
P
V^jj(ri,r2,.:,rp) = XI* - r„+i)^ + pK(r„+i) (2.19)
The (*) on the product and the summation indicate that rp+i = rj. The partition
function of Eq. ( 2.18) is similar to a classical partition function. The first term of
23
the effective potential,Vg//, originates from the kinetic energy of a particle. We may
interprete this term as a harmonic type interaction between the nearest neighbor beads
in the necklace. The coupling constant is Ci = Pmlh^lS^. In the high temperature limit,
the necklace of P beads collapses to a single point so that a quantum particle becomes a
classical particle. The classical partition function is valid in the limit e = ^/P —»• 0. The
classical isomorphism is therefore more accurate at high temperature, T, and for a large
number of P. At low temperature, the quantum particle possesses some spatial extent
associated with its deBroglie wavelength.
2.1.2 Systems of interacting particles obeying MaxweU-Boltzmann statistics
In the previous section, we discussed the thermal density matrix and the partition func
tion of a single particle in a canonical ensemble. Here, we will generalize them to systems
containing many quantum particles. We will attach particular attention to the contrast
between the discretized path integral form of the partition function of particles obeying
Maxwell-Boltzmann, Bose-Einstein, and Fermi-Dirac statistics.
The partition function for a iV-body system may be written as
Z = J dri - dr^fp{ri , - - ,r! \ !; l3) , (2.20)
where the thermal density matrix of N distinguishable particles, that are obeying Maxwell-
Boltzmann statistics, is defined by
p(ri, • • •,riv; /3) =<ri • • • ta^I \ri • • (2.21)
with |ri • • • ri\f> being an eigenstate of H. We will assume that the Hamiltonian of the
iY-particle system takes the form
iV -2 iV . jV » = E£ + E'>M + 5E«« (2-22)
«=1 t=l t>j
with Vij = v{ri — Vj) . 0(r,) is an external potential on the z-th particle and t>,j is a pair
potential between particles i and j. In order to evaluate the partition function, let us
24
discretize the density matrix by inserting (P — 1) intermediate states for each particle.
With the property of completeness of the eigenstates |ri of H. the density matrix
becomes
<ri • - -r/^l |ri • • • r^v> = <r\ • • • r;v| • • -e '^^ |ri • • -r,v>
= f n > . (2.23) •' j/=l n=l ifc=l
In Eq. ( 2.23), the subscript and superscript of denote the fcth element (or called
bead) of the necklace of the I'th electron. For the sake of convenience, we ^viH use a
new notation,]i2>= |ri and >= • •-rly' >. Using the Trotter
approximation and performing the potential operator, we may write the infinitesimal
density matrix as
= S-E." . > exp |-< '<•('•!") + E "("-I" - >•!") j I (2-24)
The term </?W| |fl(*+i) > in Eq.( 2.24) is a free particle propagator of N
distinguishable particles. If we use the result of the free particle propagator for a single
particle, Eq. ( 2.16), we have
rf) ><pS*^ . - -PS;)! e-'E.=. & |/2(^+M iV
, lt:\ .^llc\, fJtl fjfc^ fit) (k\.
n=l =/n
n=l
= /n • • -pSv" ><p1" • •
If we repeat the evaluation of an infinitesimal density matrix over all intermediate states,
the partition function of the A'^-body system can be written as
/r' \ 3NP/2 . P iV ^« (t) / n n (2.26)
1=1 i^i
25
where ^ P N
(2.27) ifc=i t=i ifc=i t=i
and
E ^ E i; "(••!" - r™). * 1 I .• 1 * f__. -<r=l i=l fc=l i>j
(2.28)
The effective potential, Vi + V2, of the N-body system in absence of quantum exchange
is similar to the effective potential of one-quantum particle system. Vi represents the
harmonic potential which corresponds to interactions between the first neighbors in the
closed necklaces. V2 is nothing but the potential energy resulting from the exLernal field
and the particle/particle interactions.
2.1.3 Two-Electron system
In the previous section, we have established an isomorphism between a classical par
tition function and a path integral of iV distinguishble particles. To extend this to
many-fermion (and many-boson) systems, we first investigate a system of two electrons.
Subsequently, we will generalize the system to iV indistinguishable fermions (or bosons).
Since identical particles can not occupy the same state by the Pauli exclusion
principle, the total wave function of a two-electron system should be antisymmetric upon
exchange between electrons. Using this fact, we will introduce a new space in order to
represent a state of two indistinguishable fermions. The new space is defined as
with one particle in state ri and one particle in state r2. The closure relation of this
space is
(|rir2> -|r2ri>. (2.29)
^ Jdridr2 |rir2}{rir2| (2.30)
The density matrix of 2-electron system can be written as
p(ri ,r2; r'i,r^;/3) = {ri,r2|e (2.31)
26
If we consider an intermediate state jr'jrj} with the closure relation, we have a convo
lution relation for two identical particles such that
P(ri,r2; = J dr"dr2 p{ri ,r2 ; r",r2; (3/2) p{r ' ( ,r2; r ' i ,r2; 012). (2.32)
We recall that the partition function is the trace of the density matrix and
Z = Jdridr2p{ri ,r2; ri ,r2;f3) (2.33)
Using the convolution relation,Eq.( 2.32), for (P — 1) intermediate states in imaginary
time, the partition function now becomes
Z = /n n ('1". 4" : "; <) (2-34) •' l/=l fc=l
where r,- = with i =1, 2, that is, each electron forms a closed necklace
with P nodes. Let us evaluate an infinitesimal density matrix with Eq.( 2.29).
= 5 (<rfV<"|e-'»|rS'+'Vf+"> + -
To evaluate the last terms of Eq.( 2.35) including the cross terms between particles, we
will consider a general case.
=<rl'lr['l| (2.36)
and the term in Eq( 2.36) is a density matrix of
a free particle propagators, which can be evaluate exactly in the same way as for the
classical particles obeying Maxwell-Boltzman statistics. With the results of Appendix
A, we have
(2.37)
27
Eq.( 2.37) shows that
<rWr[""| (2.38)
If we consider a symmetric potential^such as a pair-wise additive central potential.
Eq.( 2.35) can be simplified as
(2.39)
From Eq.( 2.34- 2.39), the partition function of two electrons finally becomes
\2NPI2 c P , r ^ = (^) / n 11^ ^ ' ' ' ' ^
^ ^ ^ „=i fc=i I
—i3-L o pV^ (2.40)
The exchange process between the two electrons is shown in the figure (2.2). By factoring
o— k k
t-
Elactrani k^l ElictronZ
Figure 2.2: Exchange process between two electrons.
the first term out of { } in Eq.( 2.40), we have
X 11 "**2*^'' )^) + )2 j
28
For convenience, we may rewrite this term by considering the product terms as
Ilf } = >+!'•'-'•= 'i . fc=l k=l k=l
= exp {-aCo t, E (>•!" - • n let (e<'-'+'I) (2.41) \ 1=1 k=l / fc=l
with
-0Co ( 1 e ^ '
-3Co ((r^*' -rj*"*"'' )2+(r^*' ; \ / 1
det (£(<^-*=+1)) =
From Eq. ( 2.40 and 2.41), the partition function of two-electron system becomes
^ 1/=! \ t=l Jt=l /
X n <!« (£<*•'+'1)6-'^=""' (2.42) Jc=l
We note that an infinitesimal density matrix of two electrons between the imagi
nary times k and k-i-l is propotional to det so that the sign of the infinitesimal
density matrix can be either a positive number or a negative number. This is the origin
of the sign problem in the simulation of many-fermion system. In particular, the parti
tion function (2.42) is an integration of terms which can be either positive or negative so
one can think of the partition function as the difference between a large positive num
ber and a large negative number to give an overall positive value. Therefore the noise
level of the partition function is large and this is the difficulty to overcome during the
simulation of a many-fermion system. Furthermore since det can be negative,
it is not possible to rewrite Z in the form of a classical partition function. Thus one
cannot establish yet an isomorphism between the fermion system and a classical one. In
order to solve the sign problem, we will adopt the fixed-node path integral method in
our calculation. The details of the fixed-node path integral method and its usage in our
simulations of many-electron system will be discussed in section 2.3.
29
2.1.4 Many-Electron system
We may extend the partition function of the two-electron system, equation (2.42), to
the partition function of a indistinguishable /V-body quantum system. Because of the
indistinguishability between particles, we will introduce an orthonormaJ basis of the
indistinguishable iV-body fermionic system:
|rir2---riv} = J^(-l)«'|rpirp2 • • •rpiv> (2.43)
where \r1r2 is an orthonomal basis of an distinguishable iV-body system, and p
is the parity of the permutation and (-1)" become +1 and -1 for even and odd number
of the permutations between fermions, respectively. The closure relation of the new basis
becomes
1 ^ • • •r^}{rir2 • • - Vtv] = 1- (2-44)
1=1
With the new basis, we may define the density matrix of the indistinguishable iV-body
fermionic system as
p{R,Ii';3) = {R\e^"\R'}, (2.45)
or
p{R,R';0) = (2.46) p
where |/Z} = |rir2---riv} and |/2> = |rir2 • • - r .v >• If we consider P intermediate
states and the closure relation, (2.44), we may have the following convolution relation
for the density matrix:
p{R,R';0)= f (2.47) . = ! where = p{R^^\ e), R^°^ = R, R^^^ = R', and e = I3IP, which is the infinites
imal imaginary time interval.
30
The partition function, which is the trace of the density matrix, for an indistin
guishable iV-body quantum system may be written in the form
Z = J dRp{R, R: 13)
= [{[dR^^^ (2.48) t=0
In order to calculate the partition function, we have to evaluate the infinitesimal density
matrix, p(R^^K t). For further development, we will assume that the Hamiltonian
of the system is H = t + Vjt where again
.V -2
t=i
i=l i.J
(i>{ri) is an external potential at ri and y(|r, — rj|) is a pair potential between particles
i and j. With the Trotter's approximation, the infinitesimal density matrix becomes
P
P
If we apply the closure relation,/dR^''^R^''^><R^''^ = 1, and the orthonormal property
of the basis >, the infinitesimal density matrix can be written as
= po{R^''\R^''-^^^;e) (2.49)
where the infinitesimal density matrix of free particle is defined as
p
The infinitesimal density matrix of free particle contains the exchange procceses between
identical particles and it can be written in a determinant form. Figure 2.3 From the
equation (2.25), the infinitesimal density matrix of free particles becomes
(-C.B'-S' .
31
I [
ele 1
(k+l) 1
I I
j
\
1
ele 2 ele3
t (
I
i
i I
(k+l) (k+I)
Figure 2.3: Three-cycle exchange process among three identicle particles. Solid lines and dashed lines represent two different exchange processes.
32
with C2 = m/2eh^. If we perform the summation with the permutation operator p, we
can simplify the density matrix by using a determinant:
/ r \ NPii e) = det , (2.50)
where the matrix is defined as
Furthermore, the exchange processes between particles can be illustrated by writing the
infinitesimal density matrix of free particle (3-particle case) as
y ^ k 3 N P / 2 e) = (^2l^y ^0 • (1 — f i2 — f23 — hi +ff23i +^312)1 (2-51)
where
1 - C 2 f ( r " = ' - r ' ^ + " V - C 2 f ( r ' ' " - r ' ' = + " ) ' - c ^ V Qiji = —.e ' J € y ' ' ' e V ' • / . (2.52)
/./s and giji's are the exchange process between two particles (t,j) and the exchange
process among three particles respectively. We show the exchange between iden
tical particles in fig. (2.3). The determinant of the density matrix of the free particles
in absence of quantum exchange is factored out of equation (2.50);
det = IJ • det ^ (2.53) t=i "
where all the exchange effects (including the sign of the density matrix) are included
in det(£'(*^'*+^') which elements are defined as (£'^^'^"''^^),j = (A^^'^'^"''^'),j/(j4^^*'^"'"^'),',.
Specifically, the matrix element of det(£;'^^''^"^^) for AT-particle system is given by
=exp [-5C. {(r!'l - (rf (2.54)
33
From equation (2.48,2.50,2.53,2.54), we finally write the partition function of the N-
fermion system as
/ r- \ 3NP/2 r P / . V P , \ = (^) / n n ''-i" -p (-i co E E (-1" - -1' ")) ^ •' \ i=ik=i J
X JJ det (2.55) fc=i
where the classical potential energy Vi is the same with equation (2.28), i.e.
= 4 E Z «>(••!") + 4 E E "(--l" - --f)- (2-36) fc=l i=l k=l i>j
2.2 Path Integral with Non-local Exchange Using the Mean Field Approx
imation
In the preceding sections, we have developed a local non-interacting density matrix which
does not describe electron correlation, since the free particle density matrix has been
obtained by using a complete set of states represented by Slater determinants of plane
waves and Slater determinant of plane waves are solutions to the Hartree-Fock equation
for free electrons. Although the local non-interacting density matrix does not include
electron correlation, in the limit of high temperature, its nodes approximate reasonably
well those of the exact density matrix[15].
We now construct an approximate form for the density matrix that includes
electron correlation. In order to treat the correlation between like-spin electrons, Hall
has proposed a non-local exchange pseudopotential[34, 35]. In the local form of the
density matrix (Eq. (2.51)), det(£'(*^'^+^)) includes ail the exchange effects. Although
the exchange occurs only between consecutive beaids in imaginary time. Furthermore, in
the limit of 6 —f 0, the matrix converges to the identity matrix and the system
collapses into a bosonic state. To avoid this undesirable behavior and inspired by the
consideration of quantum chemistry simulation, where it is well known that exchange is
a non-local interax:tion in space, Hall has suggested a non-local form of a density matrix
34
of two-electron system as follows;
P -»• n <let(£'i*'-')), (2.57)
1=1
where the matrix element (£'i*'''),j is definfed as
= exp (-^ Krfl - rl'V - (rl" - ri")^|). (2.38)
In the preceding relations, the superscripts and the subscripts label the beads and elec
trons, respectively. We note that a in Eq. (2.57,and 2.58) is a system dependent pa
rameter and the absolute value of the argument of the exponential prevents negative
weights. Because it is not easy to find a proper parameter a for a system and it is clear
that one underestimates the contribution from negative values by choosing the absolute
value, we generalize the non-local density matrix of an jiV-fermion system by choosing
the following;
det (.4(^-^+1)) n ,, • n^et . (2.59) i=l " 1=1
The non-local form of the infinitesimal density matrix of an iV-fermion system and the
corresponding partition function now becomes
, r \3NPI2 S P ,p det (2.60)
and
/ r> \ zNPn . P N / N P , \
^ u=lj=l \ i=lJc=l /
x f [ f l d e t (2.61) /=l *r=l
The non-local form for the density matrix cannot be obtained from simple Slater de
terminants of plane waves. Equation (2.60) should therefore represent electrons beyond
the Hartree-Fock approximation. A non-local density matrix would account for some
electron correlation. In Fig. 2.4, we illustrate both the local exchange process in (a) and
35
ele 1
ele 2
5 . . . P 1 4 3 2
(b) ele 1
ele 2
1 2 3 4 5 P
Figure 2.4: Local(a) and non-locai(b) exchange processes between two electrons. The number labels indicate the imaginary times.
the non-local exchange process in (b). In the non-local exchange model, a bead of an
electron interacts with all beads of any other electron.
Although the non-local form of the density matrix does not collapse to a bosonic
state when P —> oo, it still has the sign problem, because the determinant vaiues are
either positive and negative. In the next section, we will discuss and present a solution
to the fermionic sign problem by introducing the restricted path-integral Monte Carlo
method. This method has been widely used in the Monte Carlo simulation but never
used in the molecular dynamics.
2.3 Restricted Path Integral Method
There is a fundamentai difficulty in the simulation of many-body fermionic system, called
the sign problem. The sign problem arises from permutations between identical particles.
36
The contribution from even permutaions is almost the same as the contribution from odd
permutations. In practical calculation of thermodynamic properties, one can not expect
accurate results, because of large signal to noise level. The study of the sign problem
and the search for better conditioned simulation algorithm are widely discussed subjects
in the simulation of many-electron systems [7, 2, 14, 15, 53. 90].
In more recent studies, a restricted fixed-node path integral approximation has
been suggested to solve the sign problem on a many-fermion system with path integral
Monte Carlo simulation. In this approximation, the paths of all fermions in time( imag
inary time) are restriced to remain within the region of phase space where the density
matrix is positive. This approximation has been applied to liquid ^He above 1A'[15]
and the hydrogen plasma at high temperature and reasonable agreement to the existing
theories has been found. On the basis of these two studies, we may be able to restrict
phase space of electrons to the region where only positive density matrices are allowed.
The main idea of the restricted fixed-node is originated by Metropolis and
Ulam[52, 87], who have suggested the extansion of the random-walk process typically
used to simulate the diffusion equation for solving Schrodinger equation. Anderson[7] has
applied the restricted fixed-node scheme to obtain the ground state of simple quantum
molecular systems. .Anderson also solves Schrodinger equation using the random-walk
methods. We will summarize the restricted fixed-node path integral idea suggested by
Ceperley[14, 15, 16] and then apply it to the simulation of fermion systems by path-
integral molecular dynamics.
where C = C{x,t) is a concentration, D is a diffusion coefficient and i is a real time.
The solution of the diffusion equation without any restriction is
if we assume that the initial condition is C{x,t) = S{x — iq). As we mentioned in
A diffusion equation in one dimension can be written as
(2.62)
y/AnDt (2.63)
37
the section 2.1, the diffusion equation is isomorphic to the imaginary time-dependent
Schrodinger equation of a free particle density matrix:
dp{x,^) h} aV(x,/3) d0 2m dx^
by replacing i3 by it/h.
(2.64)
If we assume that the boundary condition of the diflfusion equation,
C(x-z') = 0, (2.65)
then the solution of the equation ( 2.62) becomes
1 Ix—zn—x')^ 1 (x+xn—i')^
The solution ( 2.66) has an antisymmetric form in space about x' and we can
impose an infinite potential barrier at the boundary x, which will be called the fixed
node, without changing the solution because the solution of equation ( 2.62) is uniquely
determined by the boundary condition. By virture of the isomorphism between the dif
fusion equation and the imaginary-time Schrodinger equation, we may apply the infinite
potential barrier at the node of the density matrix, with the fact that the trace of the
density matrix should be a positive real physical quantity and be spatiaJly antisymmetric
if exchange process occurs between like-spin electrons. In the path integral, the trace
of a density matrix p{x,x;(3) is always positive, but the infinitesimal density matrix
p{xi, x,+i; T = (3/P), where i = 1 to P, can be either positive or negative, we can choose
a configuration of beads of a necklace such that x, and x,+i are on the same side of the
boundary node so that p(i,-, x,+i; r) for all i are positive. In the fixed-node Path Integral
Monte Carlo method, one begins the simulation with a trial density matrix, whose nodes
are known. So if the trial density matrix is exact, the method then becomes exact.
In contrast to the Monte Carlo method, we start the simulation with arbitrary
configuration of beads of electrons without the knoweldge of the nodes. We choose
only the configurations whose determinants are positive so that all infinitesimal density
matrices have positive values. Then we calculate the forces ( called exchange forces) of
38
the configurations to generate a new configurations. More details will be given in the
next chapter.
To see the meaning of the fixed-node and to understand the sign problem more
clearly, we consider the special case of a two-body system in 3-dimensions with an ex
change process between a pair of beads (i) and (j). An element of infinitesimal density
matrix is
where and are the coordinates of electron 1 and 2, respectively. Then the
determinant of the matrix is
12 = J g-Ci {(x<'' -y'-'' -x<J> )2}
g-Ci -y'-*' )^} i
With relative coordinates x = — y^'^) /y/2ci and y = — y^-'h/V2ci, the deter
minant can be rewritten as
det(E('-'>) = 1 - e-^ y (2.67)
In the relative coordinates, the sign of the determinant now has the same sign with the
dot product (x • y). The boundary between the two regions is x • y = 0. The figure 2.5
shows these regions clearly.
2.4 Classical Isomorphism for Many-body fermionic system
In the numerical simulation of the Many-body fermionic system, we are not evaluating
the partition function directly. Instead we are trying to establish an isomorphism between
the quantum partition function and a classical partition function. With the restricted
path-integral method, we can always constrain the configurations of the particles in the
system to regions of phase space with positive density matrix. We therefore rewrite the
39
«+/
Figure 2.5: Phase diagram of the sign of determinant. The diagonal line is ® • y = 0. All unfilled dots correspond to positive density matrixes and the filled dot has a density matrix with a negative value.
40
partition function of the many-body fermionic system, Eq.(2.61), in a form isomorphic
to a classical one:
where the integration is limited to configurations with det(£'''^*'^) > 0. In equation (2.68).
V2 is a classical potential energy funtion of the position, which can be written as
V2 = V2(rj^^), (2.69)
where = |r-*' — denotes the distance between the Arth beads of the electrons i
and J. Since the ions in metal system will be treated in a classical manner, the distance
btween the ion / and he Arth beads of the electrons J is = |r/ —Ve// in equation
(2.68) is a quantum effective potential energy. The classical potential function includes
electron/electron Coulomb interactions or electron/ion interactions. In contrast, v;//
includes quantum exchange energy between electrons. From equation (2.68,2.61),we can
define the effective potential,Kg//, as
y.// = y.fr+K?/'- (2.70)
where xz-harm _ P ^ /•« -.1 ^ ) (2.11)
and
K."/' = £ E (d«(e"-'>)) , (2.72) ^ fe=l /=1 ^
where det(E^'^''') > 0 and P" is the effective number of paths with det(E^^''') > 0.
Y^arm jg non-exchange harmonic potential and is a non-local quantum exchange
potential. In the non-local form of path-integral dynamics, a quantum particle is still
represented by a necklace of P beads such that a point in the necklace interacts with
its next consecutive neighbor along the chained necklace through a harmornic potential
with a strength mP/20^h^. In contrast to the harmonic potential, the exchanges between
different particles are not limited to the nearest neighbors along the necklaces, but act
over all beads of the different necklaces.
41
In order to represent with a more practical form for implementation of
a restricted path-integral molecular dynamics, we introduce a step function d"*". The
function 9^i ensures the path restriction by taking on the values I and 0 for paths
with positive and negative det(£'^*'''), respectively. In addition, for a system containing
electrons with two diflFerent types of spins (i.e. spin-up and spin-down electrons), we may
use the fact that the density matrix is approximated as the product of two determinants
taking the form of equation (2.72); one determinant for the electrons with one type of
spin and another determinant for the electrons with the other type of spin[37]. We now
rewrite the as
1 down PP.
n?;" = -3EEE^I'' (detCEC'l)) «5,. (2.73) ^ «=«p it=l 1=1 '
where
p:=j:{:oL-k=i 1=1
In the previous two equations, s in the summation denotes the spin of the electrons.
Up to now, we have written the partition function of a quantum fermion system
in the form of a classical partition function. In the following section, we will use molecular
dynamics to sample the newly developed classical-like effective potential for a quantum
system.
2.5 Moleculeur Dynamics
2.5.1 Background
If one has a microscopically well-defined physical system, one can use Molecular Dy-
namics(MD) method to calculate the physical properties of the system. MD method
computes phase space trajectories of a system of many particles which individually obey
the classical Newtonian equations of motion. Specifically, if a iV-particle system is de-
scribed by a classical Hamitonian H = + Et>i where o,- is a velocity of a
42
particle i and (pinj) is a pairwise centeral potential between particles i and j separated
by Pij = |r,- — Tjl, one can anaijrtically specify the phase space trajectories, (r,(t),
by solving the Newtonian equations of motion with a certain initial conditions. In other
words, we can find the time evolution trajectories by solving the following equations;
. .9,
and ^ at
where Fi is the total force on the particle i, which is the sum of the forces on a particle
i from all the other paticles in the system, and also the force on the particle i from a
particle j, can be obtained from f^j = —In order to solve the equations
of motion in a numerical MD simulation, one discretizes the differential equations of
motion. Several numerical schemes are then used to integrate the equations of motion.
Among these schemes a finite difference method is often used.
Since Alder and Wainwright[4] used MD method for N-body system, the MD
method has been developed and applied to simulate a large variety systems[10-24]. The
main advantage of the MD method over the Monte Carlo method is that it allows the
calculation of time-dependent properties in addition to the equilibrium properties which
can be obtained by either methods.
Early simulations were carried out for systems where the energy was a constant of
motion[4, 5, 70, 86]. Accordingly, properties were calculated in the microcanonical en
semble where the number of particles, the volume, and the energy were constants. The
other model of the MD method is the constant pressure method. This method was intro
duced by Andersen[6] and subsequently generalized by Parrinello and Rahman[62, 63, 56].
The volume of the system is treated as an additional dynamical variable in this method.
The MD method of the constant pressure is applied to structural changes in the solid
state. However, in most situations one is interested in the behavior of system at constant
temperature. This is partly due to the fact that the appropriate ensemble for certain
quantities is not the microcanonical but the canonical ensemble. After W6odcock[93]
used a constant temperature MD with a momenturm rescaling procedure, in which the
43
velocities of the particles are scaled at each time step to maintain the total kinetic
energy at a constant value, several constant temperature MD methods have been pro
posed. Haile et a/.[32] have examined analytically the constraint method based on the
momentum scaling procedure. Andersen[6, 48] proposed a hybrid of MD and Monte
Carlo method and introduced stochastic collisons in the phase space trajectories. .\n-
other constant temperature MD Method has been proposed by Hoover et a/.[41. 45]. In
this method, the force on a particle is constrainted such that the total kinetic energy is
constant. Introducing virtual variables, Nose[56, 57] has generalized the constant tem
perature MD method.
In our MD simulations, we adopt the constant temperature MD method based on the
monentum rescaling procedure. In this method, the equilibrium distribution function
deviates from the canonical distribution by order of N~^ for a N particles system [57].
Thus the average quantities calculated with this method will be in error of C?(iV~'). In
the momenturm rescaling constant temperature method, the momenta of the particles
in a simulation cell are rescaled at each time step to maintain the total kinetic energy
at a constant value. If T/ is an instantaneous temperature of the system,
where KE is the total kinetic energy of the cell and p,- is the momentum of the Jth
particle. Using this relation, the momentum can be rescaled such that
In molecular dynamics simulations, we generally restrict our particles within a
physical volume, the basic cell or the MD cell, to retain a constant number density of
the particles of the system. Let the system consist of N particles within a cubic basic
cell which an edge length , L, and a volume ft = When the system evolvs in time,
KE
(2.74)
,scaled
where yJTrej/Ti is the scaling factor and Tref is the desired temperature of the system.
44
particles will hit the surfaces of the cell and be reflected back into the cell. Especially
for the system with a small number of particles, we would have some unexpected effects
from the restricting surfaces. In order to reduce the surface effect we impose periodic
boundary conditions(PBC)[80]. With PBC, the MD-cell (or the basic simulation cell) is
copied an infinite number of times by identical image cells. We may, therefore, write a
physical quantity A(r) in the MD-cell as
A(r) = A(r - \ - n • L ) (2.76)
where n = (wi, nj,,nj) whose components are integers, L = { L x , L y , L z ) is the size of
the simulation cell, and r is confined within the cell, i.e. [r| < L. Equation (2.76) means
that if a particle crosses a boundary of the cell, it re-enters through the opposite side of
the cell at the same instant. Figure 2.6 shows the behaviors of particles with the periodic
boundary conditions in two dimensions. There are 26 image cells in three dimensions.
With the periodic boundary conditions, the finite MD-cell is extended to infinite by
indentical image cells. In other words, due to the PBC the potential energy is represented
by
^ H L \ ) , (2.77) i < j n i < j
where $(r,j) is a pair potential, = |r,- — rj|, and r, and rj are restricted with in
the MD-cell. In order to avoid the infinite summation in the last term in Eq. (2.77),
we introduce a cut-off range (r^) for the potential[92, 40]. .A. particle in the basic ceil
interacts only with each of the N - I other particles in the MD-cell or its image cells.
Effectively we may cut off the potential at a range
Tc < L/2. (2.78)
With this cut-off range rc, we rewrite the potential energy as
^ = 1 2 E - • = ) ) ( 2 r a ) t (< j ) j^MDcell
In figure 2.6, we illustrate the cut-off range of the Arth particle by a dotted circle. We as
sume that the jth and the fcth particles remain in the same positions but the ith particle
45
O k O k C k
C- ..
i O j i'
^O-
i o j
r-i o ••
i'
Ok ^k O k
0 . 1 O
J '' i
j i-
.O:
i O
j i'
o'' o O k
a
i ••
j i* i "O
j i-
JD- . . i • o
j i'
Figure 2.6: Basic simulation cell(thick solid line) and its 8 image cells in two dimensional system with PBS. The size of the basic cel l is L. The radius of the dotted circle is L/2. The dotted circle area is the effective range of interaction.
46
moves from a site i to another site i' after some time At. At time t, the Arth particle was
affected by the ith particle in its image cell, not the ith particle in the MD-cell. However,
potential between the ith and the Arth particles is calculated within the MD-cell at time
t + At.
Although we can effectively simulate a system with MD method with PBC and a trun
cated potential, such a truncation may have a determinential effect on systems with the
long range interactions such as systems of particles interacting via the Coulomb poten-
tiai. There is a very elegant procedure, known as Ewald's method, to calculate effectively
Coulomb potentials with PBC[43, 71, 94]. The Ewald's methods will discuss in section
2.5.2 Restricted Path Integral Molecular Dynamics Method
In this section, we develop a classical Hamiltonian useful with molecular dynamics sim
ulations based on the restricted path-integral representation of quantum particles. Once
we set up the Hamiltonian of a system, we can calculate the phase space trajectories of
all particles with the molecular dynamics method.
Let the system consist of Nei unpolarized electrons. For the system, the num
ber of spin-up electrons is the same as the number of spin-down electrons. If there is
no external force, the total potential for the system is the sum of the electron-electron
Coulomb potential and the effective potentials given by equation (2.71,2.73). By con
sidering a kinetic energy term for the N^i electrons, the classical Hamiltonian can be
written as
3.2.
"el r- r i ,2 i
1=1 k=l k=l i{>j) j=l |r, -, down P P 1
(2.80)
where P P
k=l 1=1
47
In Eq. (2.80), m" is an arbitrary niass of a bead (we chose m" = 1 a.m.u.)[64] used to
define an artificial kinetic energy for the quantum states in order to explore the effective
potential surface, Vejj, and rric is an electron mass. We recall that the subscript and
the superscript on the position are for labelling an electron and a bead in this electron,
respectively. The second term in the above equation accounts for the electron-electron
Coulomb potential energy. The forces derived from the non-local exchange potential,the
last term of Eq. (2.80), are calculated as means over the restricted paths with positive
determinants. Therefore, an effective force calculation requires a satisfactory sample
of such paths. Since the exchange potential offer a barrier to paths with negative de
terminants, it biases the sampling of phase space toward configurations with positive
determinants. Although many configurations with negative determinants may exist and
evolve, they do not contribute to the exchange forces.
To extend the restricted PIMD method to cdkali metal system, we simply add
the potential terms related to the classical ionic degrees of freedom to the Hamiltonian
of the electron plasma system. For a system containing Nei unpolarized electrons and
^ion ions, the Hamiltonian becomes
^ton 1 ^lon ^ton I H = H.,.+ £
1>{J) J
^ton ^ton t + E E E p^pWo(«/ - r!'^^), (2.81)
/=l 1=1 /=l ^
where M/ is the ion mass and R is the position of an ion. In Eq. (2.81), Heu is the
Hamiltonian of the electron plasma described in Eq. (2.80) and the second term is the
kinetic energy of the ions. ^(R[j) and V^seurfo are the ion-ion potential and the ion-
electron pseudo potential, respectively. These two potential energies will be discussed in
section 3.1.
48
2.6 Physical Quantities
2.6.1 Average of Physical Quantities
Using the molecular dynamics method, one can obtain the phase spcice trajectories,
( r(i), p{t)), and subsequently any microscopic physical quantites, say A{r{t),p{t)), as
functions of time. If the system is ergodic, we may compute the time average of A{t) to
find its corresponding macroscopic property. In other words, the average of the physical
quantity A can be calculated as
2.6.2 Energy Estimator
We have constructed the canonical partition function for a many-electron system with
P intermediate states in the imaginary time. To evaluate the mean energy, let's rewrite
the partition function (Eq. (2.68) ~ (2.73)) for the convenience:
where the potential energy Vj includes all classical potentials such as the short range
interatomic potential and the pseudopotentiaJ between an electron and an ion as well as
the long range coulombic potential. The non-local quantum effective potential, is
given as
/ A { r { t ) , p { t ) ) d t . J o
(2.82)
where down P P
Now we can evaluate the energy estirmator by
(2.83)
49
The estimator for large enough P becomes
IE\ - /V / /V p
(EEt-: \ j= l Ar=l
)') + (^(/'KTf)) + <^2> • (2.84)
In Eq. (2.84), the angled brackets denote ensemble average and these quantities will
be evaluated by time average calculation, Eq. (2.82). Since the last term of the energy
estimator equation can be interpreted as a potential energy estimator, the kinetic energy
estimator, (KE), becomes
The first term in the left-hand side of Eq. (2.85) is a constant for given P, iV, and T. The
second term of the equation is the harmonic eiFective potential contribution to the the
kinetic energy estimator. Both terms are easily calculated with minor computing time
if all configurations are known. However, the third term in the equation, i.e. {KEexch}^
which is the exchange potential contribution, cannot be calculated unless we pay a major
computing cost, because the number of required operations is the order of ~ iV^P^. We
will give a full derivation (KEexch) term in ."VPPENDIX 2 and simply write the result
here:
(2.85)
where (KEexch) is defined by
j n down P P , \ = (^EEi :^Ndet (E( '=- ' ) )0L)
\ a=«P fc=l /= l ^ /
(2.86)
(KEexch.) — 2^ 2->
down P P (2.87)
3=up k=l 1=1
where is a.n N x N matrix and its element (^i*^''^)at is given by ik,l)
(2.88)
and
e " = ( rW-r ' " )=-( r (" - r ( ' )p .
50
2.6.3 Correlation Functions
Pair Correlation Function
An important physical quantity that we can easily compute during a simulation is the
pair correlation function (also called the radial distribution function), g{r), which tells
us the actual spatial distribution of one kind of particles with respect to another kind
of particles. With the pair correlation function, we can compare simulation results with
the experimental structural data, like that obtained from X-ray diffraction. The pair
correlation function between the (i) and {j) types of particles ,^,_j(r), is defined as
where Q is the volume of a simulation cell, N j is the number of the type j particles, and
n,(r) the number of the type i particles situated at a distance r and r + <5r from a type
j particle.
Time-Correlation Functions
We often calculate the time-correlation functions to understand macroscopic transport
properties of a system. One of the time-correlation functions is the mean square dis
placement (MSD). The mean square displacement is defined as
rT M S D { T ) = lim / (r(«-I-r) - r( r ) ) ) 2 (2.90)
T->oo Jo
with r being a correlation time and r being the position of a particle. By calculating MSD
from the trajectories of particles, we may determine whether the processes developed by
a system is a diffusion process or a vibrationaJ one. For normai diffusion processes, i.e.
one that occurs slowly with respect to microscopic times and for which spatial variations
are smooth, we have
D = ^-MSD{t), (2.91)
51
where D is the coefficient of self-diffusion[51]. The mean square displacement is a linear
function of time for a liquid.
Another time-correlation function is the normalized velocity autocorrelation
function (NVAF), which is defined as
where the velocity autocorrelation function (VAF) is given by
VAF(r)= Um / (t»(t + r) - tj(r)))2. (2.93) T- oo Jo
V in Eq. (2.93) is the velocity of a particle. Using the velocity autocorrelation function,
we can also calculate the power spectrum or spectral density G{f) of the NVAF. By the
Wiener-Khintchine theorem[42], the power spectrum is the Fourier cosine transform of
NVAF, i.e.
G(/) = lim 4 [ NVAF(r) cos(27r/r) dr, (2.94) T-+00 Jo
where / is the frequency of a vibrating particle. The power spectrum is nothing but the
vibrational density of states of a collective ionic motion.
52
CHAPTER 3
PRACTICAL IMPLEMENTATION
3.1 Treatment of Classical Potential Energy
We apply the restricted quantum MD method to both an electron plasma system and
an alkali metal system. In this section, we discuss the classical potential energy for each
system. Electon-electron, electron-ion and ion-ion interactions include the long-range
Coulomb potential. We therefore introduce the Ewald summation method to calculate
the a long range potential in a small basic simulation cell with periodic boundary con
ditions A Born-Mayer ion-ion potential and a ion-electron pseudopotential will be also
introduced for the metal system.
3.1.1 Electron Plasma System
The electron plasma system consists of N unpolarized electrons with an uniform posi
tively charged background. The positive background is imposed to neutralize the total
charge of the system. In path-integral MD, each electron is considered as a necklace of
P beads. The only classical potential, V2 in Ekj. (2.68), of the electron plasma system is
the Coulomb pair potential. From Eq. (2.28) and (2.49), the Coulomb potential can be
written as
(3.95)
53
where is the distance between the bead (A:) of the ith electron and
the bead (fc) of the jth electron. We illustrate the Coulomb interaction between beads
in Fig. (3.7). We note that Coulomb interaction occurs over all the pairs of beads in the
1*1
Figure 3.7: Coulomb interaction between two electrons i and j at different imainary times.
discretized necklaces. This means that the electron-electron Coulomb interaction is not
retarded in imaginary time.
In an MD simulation with the periodic boundary conditions, a selection of the
size of a basic simulation cell is limited by computing capability. .A.lso the range of
an interaction between two particles is restricted within half the size of the cell (i.e.
Tc < 1/2L). Because of these restrictions, we may neglect a critical amount of long
range interactions, such as the usual slow converging Coulomb interation, between a
particle in the basic cell and a particle in any of the image cells. There is, however,
a very elegant and well-known model, known as the Ewald summation method[21], to
accrately account for the contribution of long range interactions to the energy. Following
the method, we consider a lattice made up of ions with positive or negative charges and
shall assume that ions have a spherically symmetrical gaussian distribution, with charge
54
density at radius r proportional to rj being the Ewald parameter. The calculation
procedure of the Ewald potential have two distinct but related parts. One is computing
the potential from a structure with a gaussian charge distribution at each ion site. The
other one is the potential of a lattice of point charges with an additional gaussian charge
distribution of opposite sign superposed upon the point charges. The parameter t] is
choosen such that both potentials at reference points converge rapidly.
The original Ewald summation method has been further developed by Nijboer et a/.[55].
They generalized the summation method to the interaction having form of 1/r". This
generalized Ewald summation method has been used in the Monte Carlo simulation of
the classical one component plasma by Brush et a/.[9] and of the fermion one component
plasma by Ceperley[l2]. Brush et al. showed that the Coulomb interaction for the
electron plasma system with PBC can be written as
+ E erfc(,n,) - " (3.96)
where m = —e for all i, 0, = L^\s the volume of a cell, fc is a wave vector , and rij=ri — rj
and rij is its magnitude. The wave vector can be written with an integer vector n as
k = ^n. The usual complementary error function erfc is
erfc(x) = 1 — erf(i) = 1 ^ f e~^ dy. yT JO
The last term in Eq. (3.96) is the contribution of the positive background. In path-
integral MD simulation, we may write the corresponding potential energy to Eq. (3.96)
as
= (3.97) fc=l £=l
where the summations are performed over all beads of the electrons in the basic simula
tion cell. Specifically, the potential energy corresponding to the Ewald summation can
55
be written as
1 / 2 f 7r |np'\ f2Tr (fc)'\
+ 5 E ( 3 . 9 8 ) «=i ^ i(^j) j=i ,=i I J
where n is a reciprocal lattice vector in units such that its components are integers. If
we use the relation the first term in the right-hand side of equation
(3.98) can be made to take a considerably more eflBcient form[71]
1st—term
With equation (3.98), the double sum over i and j in equation (3.99) converts to two
single sums, i.e. ~ N'P calculations are reduced to ~ NP calculations.
3.1.2 Alkali Metal System
The alkali metal system is composed of N ions and N electrons in a cubic basic simulation
cell. In the model, the electrons are discretized in imaginary time by P beads for each
electron and the ions are dealt in a purely classicai manner. The total electrostatic
potential energy can be divided into three parts and written as
V2 = V^-' + V^-"^ + , (3.100)
where and are the electron-electron, the electron-ion, and ion-ion
potential energies, respectively. We use the Coulomb potential for the electron-electron
interaction which has been given at Eq. (3.95).
For the ion-ion pair potential, we use the Bohn-Mayer (BM) form of potential suggested
by Fumi and Tosi[27, 81] for alkali halides to model the interaction between ions. This
56
potential includes also a Coulomb potential term. The BM potential is
A T iv-l 2
I>J J=l (3.101)
where r/j = |r/ — rj\ is the distance between an ion pair I and J, and A/j and pij are
parameters of the potential. The values of the parameters can be found at the reference
[72]. We use Au = 0.6172 xlO^° eV and pij = 0.1085 A for potassium metal. The first
term of Eq. (3.101) is the Coulomb potential and the second term of the equation is the
short range core repulsion. This short range repulsion models the interaction between
the electrons in the core of the ions.
In order to deal with the electron-ion interaction, we adopt the empty core pseudopo-
tential model[R.W. Show]. In this model, we assume that a positive ion is a conducting
sphere with radius Rc and the total charge +e. The local pseudopotential on the /th
ion from the fcth bead of an electron j is defined as
from the center of the /th ion. Fig. (3.8) shows the pseudopotential. In the figure
the p>seudopotential (a) is represented by the sum of a usual coulombic potential (d), a
repulsive p potential inside the core (b) and a constant potential inside the core (c). The
total potential on the /th ion then becomes
where d[j is equai to 1 for r/j < R^, otherwise 0. We choose the core radius as Rc =
2.22 A[38]. The electron-ion potential energy can be calculated by V(''/)-
To optimize the calculation, we do not use the Ewald summation but simply replace
the long range Coulomb potentials in and by a faster converging
-e/Rc, if < Rc
(3.102)
- e / r \ y , i f r \ y > R c
( k\ |r/ — V j I is the distance of the Arth bezid of the electron j measured where r
57
(•) (b) (e)
R e
Figure 3.8: Empty core pseudopotentiaJ model for the ion-electron interaction. The sum of the potentials (b), (c), and (d) is equal to (a).
effective potential of the form
where rj is an Ewald parameter. We choose the Ewald parameter rj = 5.741/L. With
this choice of the parameter, the reciprocal space sum in the Ewald summation, the first
term of the left-hand side of Eq. (3.96), is small compared to the real space contributions
and may therefore be neglected[48]. Thus Eq. (3.104) become close to the the Ewald
summation.
3.2 Algorithm of Molecukir Dynamics
3.2.1 Overview
In this section, we will describe the overall algorithm of path-integral molecular dynamics
(MD) for both the eletron plasma system and the Alkali metal system. The MD method
developed by Alder and Wainwright (1959) is conceptually the simpler. In our model
r r ->• -erfc(T/r), (3.104)
58
system, the classical interparticle potentials are pairwise additive and central, i.e. the
total potential energy, ^ciasaicah for a system of N particles can be written as
iV N ^ciasaical — (3.105)
«=l j>i
where 0tj(r,j) is the potential between particles i and j which are separated by r,j =
|r,- — rj|. The total force on the ith particle is then calculated by
N PiiTi) = Y. /.jM' M i = 1^2,..., N (3.106)
j(><)
and
= -V>.j(ro), (3.107)
where /,j(r,j) is the force on the particle i from the particle j. We note that the potential
in this section is not including the exchange potential that will be discussed in section
3.3 with the exchange force. In an MD simulation, we compute the trajectories of a
collection of the particles in phase space which are the numerical solution of Newtonian
equations of motion. In order to set up the numerical algorithm, we start with the
Newton's force law
^ ^ 2,N. (3.108) Clu TTh
To solve the differential form of the equation, Eq. (3.108), on a computer we apply a
finite difference scheme for the second-order differential equation. We then rewrite Eq.
(3.108) as
= ^^{r.(i + A0-2r.(f)+r.(i-A«)}, (3.109)
where At is the simulation time step. From Eq. (3.108) and (3.109), we have
r.(f + A0 = 2r.(f)-r,(f-A^) + ^^^i^(A^)^ for i= 1, 2 , . . . ,N. (3.110) m
With this equation we can obtain a new position at time t + A t from the positions at
two consecutively preceding time steps and the force acting at at time t. Starting from
59
r,(i = 0) and r,(Ai) provided as an initial conditions we can compute ail subsequent
positions by appling Eq (3.110) to the system recursively. We then calculate the velocity
of the zth particle with the Euler backward scheme as
' = 1»2, (3.111)
Equations (3.110) and (3.111) is called the simple "leap-frog" integration algorithm used
by Verlet[86].
In our simulation, the number of particles, the volume and the temperature
of the system remain constants. A pair of initial configuations, (ri(0), r,(At)), are
randomly generated but subject to the constraint on the total momentun
iV m,p, = 0, af i = 0. (3.112)
:=1
where m, is the mass of the particle i . As the system develops in time the total kinetic
energy will be fluctuating due to a re-adjustment of both particle positions and velocities,
because there is a coulping between the temperature and the total kinetic energy;
= (3.113) ^ i=l ^
where T { t ) is an instantanious temperature at time t and ks is the Boltzman constant.
In order to avoid the temperature fluctuation, we rescale the velocities, u,(f + At), of
all particles in the system with a scale factor, x- For 3- constant reference temperature
Tref, the scale factor is
*<" = (3.114)
where T { t ) is obtained from Eq. (3.113). The velocity at time t + A t \ s then rescaled as
Vi{t + A«) = x(f)i;,(i), for i = 1,1,..., N. (3.115)
Algorithm Al. Overall MD-routine
1. Specify initial configuations {r°} and {r,-}.
2. Compute the velocities at time step n as
c? = (r? - r"-i)//i
60
3. Compute potentials and forces and F" with Eq. (3.106) and (3.107).
4. Compute the advanced positions at time step n+1 as
^n+i ^ 2rf - + Ff/iVm. ; ^^.(S.llO).
5. Compute the temperature at time step n as
6. Compute the scale factor x" as
x" = y/rZr^-
7. Set rf ^ r"-' and r".
8. Rescale the positions as
r, r. + x - r - ) .
9. Repeat steps from 2 to 8.
In Algorithm Al, we summarize the above calculating processes in order. We here
substituted the time step At to h. For the simulation of an ionic system, the ion reference
temperature T"" is independent to the electron reference temperature T'jfy. Thus we
determine the scale factors independently. Since the strong bead-to-bead harmonic forces
for the system with large number of bead, P, on an electron necklace may leaxl to non-
ergodic behaviors[33], the electrons are re-scaled a different way from each other. In the
step 5 of Algorithm Al, we calculate different instant temperatures for each electron
by the the summation over the beads on the same electron. The associated scale factor
for each electron is calculated by the step 6 of the algorithm.
3.2.2 Creation of the Initicd Configurations
Every MD simulation is started with a different initial configuration obtained from ran
domly generated beads in every electron necklace. The initial bead-bead distance is
61
predetermined according to the system temperature and the expecting kinetic energy
by using the harmonic kinetic energy estimator, which is the first two terms of the left
hand-side of Eq. (2.85). We firstly generate the initial beads (r-^\ for i=l,..,N) of all
electrons with a random number generator (rani) (see the Numerical Recipes[68]. All
initial beads are generated within the basic cell. Since the electrons of our model are
unpolarized, one half (iV,) of the electrons have diflFerent spins from the other half of
the electrons. We assume that the first half and the second half of N are labeled for
the spin-up and the spin-down electrons, respectively. The following procedures are for
the spin-up electrons. Starting from the position of the first bead, about 90% of sequen-
tiai beads are generated with the predetermined constant distance in three dimensional
space. In order to reduce the time for the system to reach equilibrium from its initial
configuration, we choose that the maximum spatial extension of the necklace is equal
to 1.6 L but the necklaces remain within 0.6 L from the basic cell. The rest of the
beads are used to construct a complete closed necklace. We repeat this procedure for all
other electron necklaces and have a set of initial configuration {r|*^(t = 0)}. Another
set of configuration, {r|^'(Af)}, is created from = 0)} by generating random
displacements of the beads such that the total momentem is conserved, i.e.
r-'^'(Ai) = r-*^'(0) + for i = 1,.., and Ar = 1,.., P
with P
(3.116) Jt=i
We can use the above procedures for the spin-down electrons. Figure (3.9) shows the
initial necklace configurations for two electrons. The center square in the figure is the
basic cell. The necklaces in the figure are spreading to an image cell and they will be
broken if translational periodic boundary conditions are used. We discuss this in the
following section.
62
1-
• 'T 2STi
-r - M
Figure 3.9: Initial necklace configurations for two electrons in an electron plasma at T = 1300K and r, = 5. Only 1/3 of the beads (P = 480) denoted by filled circles are shown in xy-plain.
3.3 Periodic Boundary Conditions in Path Integral MD
The electron necklaces does not always remcdn confined in the basic simulation cell
but spread over its neighboring image cells. When some beads of an electron necklace
expand to the image cell (or cells), the continuity of the necklace is broken by the wall
(or walls) of the simulation cell. Assume that two beads belonging to the same necklace
of the electron i are at and within the simulation cell. Let us suppose that
bead (A: +1) moves outside the simulation cell in the direction (1). then its coordinates
within the simulation cell becomes — (Li,0,0), where Li is the length of the cell
in the direction (1). Periodic boundary has broken the continuity of the necklace. If we
calculate the distance between beads k and A: + 1 by the usual MD method where the
distances are calculated without maintaining the necklace continuity, we can not have
the true value of the distance between two consecutive beads. In Fig. 3.10 we show the
effect of PBC on the continuity of the necklace for P =4. We assumed that beads (1,
2, 3, and 4) make a complete necklace. If we calculate the distance between the beads
63
i
1
i 2
(a) ! 9"
i :
3 O-
<
i
1 2* V
» Q
i i
>4 ^ ' i i___i4
* L »-i
t 1
L
O
r C-1---0
I I
.__04 '
(b)
Figure 3.10: Beaxis of an electron necklace in the simulation cell and the image ceils. Beads (1 and 4) are in the simulation cell and beads (2 and 3) which belong to the same necklace are in the image cell. In the usual MD method, we use the filled circles in (a) to calculate distances between beads. In (b), the necklace is reconstructed after translating beads (2 and 3) by L from the left
64
within the simulation cell, for instance between the filled circles in (a) of the figure, the
calculated distance between 1 and 2' is not equal to the true physical distance between
beads 1 and 2. In our simulation, the position of a bead is denoted by the position vector
in the simulation cell and the address of the image cell where the bead is located in order
to allow for the reconstructing of the complete necklace. In Fig. 3.10 (b), we reconstruct
the necklace by translating beads 2 and 3 from 2' and 3' by -L, respectively. We note
that the reconstructed necklace must be counted only once in the simulation cell. If one
necklace, like the necklace of the filled circles of figure 3.10 (b), is considered as the one
in the simulation cell, then the other identical necklaces, like the necklace of the unfilled
circles of the figure, should be ciddressed to a neighbor cell.
PracticaJly, in the MD simulation, the position of the fcth bead of the ith electron is
denoted by and where is always defined inside the simulation cell and ( k )
XJ contains the information necessary for reconstructing the integral necklace of the
electron. In other words, a set k=l,2,..,P} represents a complete closed
necklace of the electron i. When we measure the distance between two beads of the same
electron i, we calculate
However if one asks about the distance between two beads in different electrons with
PBC, it does not become an obvious question. The difficulty arises because there is no
reference point for the distance measurement. In Fig. 3.11, we show the beads of two
electrons i and j, and their images. The larger circles represent the beads of the electron
i and the smaller filled dots the bead / for the electron j.
If we define a reference point as -f- then we can easily calculate the
distance from this point to the ith bead of any other electrons which can be in any of
27 cells. For a example, if we want to measure the minimum distance between the Arth
bead of the i electron and the /th bead of the j electron in any cell, we can calculate
(3.117)
- XS'^) - (rf + n„uL), (3.118)
where ticeii is an integer vector.
65
I* 1 •
elej
1 •
o
6 \ 6
P
\ 6 o
o
I • I*
ele J
. •
o
ele i
0-
I- o' elei
o
o
• 1
•
1
•
1
o
d
6
d
6
o
o
Figure 3.11: PBC diagram with two electrons. The smaller filled dots represent the /th beads of the electron j and the larger circles represent the herds of of the electron i. We assume that the beads denoted by the large filled circles are in the simulation cell.
66
3.4 Calculation of Quantum Effects
3.4.1 Evaluation of det(£'^'''')) of the Effective Exchetnge potential with PBC
When we evaluate the exchange potential and the forces, we encounter some practical
and fundamental difficulties. One of the difficulties is associated with the number of op
erations to calculate the quantum non-local exchange effective potential. Since periodic
boundary conditions axe imposed on the system, the amount of operations dramatically
increases. For a simulation in 3 dimensional space, there are actually 26 image cells
which are the exact copies of the basic simulation cell. Fortunately, both the formulas of
the effective potential and the corresponding force have symmetric forms and have nat-
ually parallelizable forms with respect to the number of beads, P. These two facts may
be used to optimize the aigorithm. First we discuss optimization of the calculation of
the exchange effective potential. The parallel computational algorithm will be discussed
in a later section.
In order to discuss the algorithm for the calculation of the effective potential with
the path integral molecular dynamics, let us rewrite the effective potential of N'-iso-spm
electron system:
. (3.119)
where
K57" = f;E^(r!''-••!»«>)= (3.120)
and
= "i E E K,- (3-121) ^ p=i,=i ^
In the above equations, and are the harmonic potential and the exchange
potential functions, respectively. A matrix element (£'^''''^),j is given by
[-^ {(.(-I - - r!-')^}] (3.122)
In the above equations, the subscript and the superscript are used for the index of
electrons and the index of node of a given electron, respectively. With the periodic
67
boundary conditions in 3 dimentions, the system is constituted of a basic simulation cell
and 26 image cells. As the usual molecular dynamics, an electron in the simulation cell
interacts with all other electrons in the simulation cell and the neighboring cells. Then
the exchange process must be considered between an electron in the simulation cell and
any electron in the neighbor cell as well as exchange process between two electrons in the
simulation cell. We, therefore, have to expand the size of matrix to (27iV' x27N')
instead of (N' x N'). For instance, if one has 27 spin-up and 27 spin-down electrons in
the simulation cell (like in our metal system), the size of the matrix will be (729 x 729).
It is practically impossible to calculate efficiently determinants of this size. To find a
good approximation, let's expand the determinant as the following:
N' N' jV iV N' det (£(-•'1) = I - Z E E f i r ' + E E E E sS"'
cells t=l j=l cells t=l j=l Jfc=l
N' N' N' N' -EEEEE' '1 ;« ' + - - (3 -123 )
cells »=1 J= l /= l
where the summation JZce/Za means that every index of particles (i.e. i, j, k and /, ..) is
running over the all 27 cells. 1 arises from the diagonal term of the determinant, because
all diagonal terms of the matrix are 1, and /,j, gijk, and hijki represent 2-cycle, 3-cycle,
and 4-cycle exchange processes, respectively. Since we are interested in calculationg
exchange force acting on the electrons within the simulation cell, we can restrict the ith
electron within the simulation cell so that the index of the tth electron in Eq. (3.123) is
independent to the summation over cells. In other words, the size of the matrix
is reduced to (27iV' x N'). To gain some intuition for this expansion, the second term of
the equation which corresponds to exchange between pairs of electrons may be written
as
E E fir-' = E E exp [-1^ {(r!" - r f f - (r!-l -cells i,j=l cells i.j=l I
+(rl-' - r!")2 - (r;»l -
N' N'
cells t=l J=l
68
,V' f iV .V N' 1 = E E + E + • • •+E [ • (3-124)
1=1 ij=i j=i j=i )
where
cp = _ r(''))2 'J > '
4 = (3.125)
L^f = 4 + 4 _ _ ^2^ R- = 1,2, • - •, 27. (3.126)
The index i in the above equations is restricted to electrons within the simulation cell,
but the index j can point to any electron over all neighbor cells including the simulation
cell itself. Figure [ ] shows two-cycle exchanges between a pair of electrons. One of the
pair of electrons should be kept within the simulation cell.By choosing a minimum value
of Llf^' over all K, we have
N' N' N' N'
EEE/r ' -EE ' - '™" cells i=l j=l i=l j=l
V \ (3.127) I Kl^Ko) )
The second term in the above equation is much smaller than I for most configura-
tions.Thus we can make the approximation:
1 ' N' iV iV' iV N' ;V' N'
E E E ft' = E E = E E E E (s-^s) cells i=l j=l i=l _7=l :=lj=lt=lj=l
Furthermore, we can contineously apply this approximation to higher cycles of exchange
processes. Then we may have that
^ 1 - E E +E E E sisri. - • • • (3129) i=l j=l t=l j=l fc=l
or
det(£J27yv'x27iV') ~ (3.130)
This approximation , which includes all possible exchange cycles, works better than the
weighted 2-cycle approxmation suggested by Hall.
We details the practical computing algorithm at fiqure 3.12.
69
AJgorithm 2 Calculaing det( *).
1. for i = 1 to P 2. for j = i to P
3. for ie = 1 to N-1 4. for je = ie to N
Calculate dl = [(r\,+X\,-Xl)-r',S-
5. for cell = 1 to 27 Calculate
rf3(ce//) = ({rU + X;. - X{^) - (r]^+n„nL)f d4{cell) = Hr), + X], - Xj.) - (rf, + n„uL)? arg(ceU) = \d3(cell) + d4{cell) -dl- d2\
6. Find the cell number. Cm, such that arg{ceU) has a minimum at the cell and set
argl = —Co(rf3(cm) — di] and arg2 = —Co(t/4(Cm) — d2) arol
cU.') — g-argt je.»e ^
7. Repeat 4 and then 3.
8. Put 1 to all diagonal elements of and Calculate det(£^''*') by using the LU-decocomposition algorithm. Evaluate det(£« *>),
if det(£'^' *') > 0, calculate exchange force and then repeat 2 if det(E<' **) < 0, repeat 2.
9. Repeat 1.
We use that Co =
Figure 3.12: Algorithm for the calculation of the determinant det(E(''-'^).
70
3.4.2 Effective Force Calculation
To perform the molecular dynamics, we have to calculate force on each particle. The
effective force originating from the effective quantum potential is given as
f e / f = fk arm "t" f exch
where fiuirm. f exch the harmonic force and the exchange force, respectively. Let's
consider the effective force on the bead (Ar) of an electron (t). The harmonic force may
be written as
{(-!" - - c!*-" - -l")} (3.131)
The harmonic force is calculated with a closed necklace as shown in Fig. 3.10 (b). The
exchange force can be derived by calculating of ^exch with the following
matrix algebra:
n
det.4 = ^aij Aij, i=l
«=l j=l
where A is an ( n x n ) square matrix, a,j is an element of matrix .4, and A,j is a cofector
of the element a,j. The exchange force on the bead (k) of an electron (i) is given as
where is a cofactor of a matrix element In odrer to solve this equation,
we explain the details at APPENDIX C. The exchange force is
= (1^) f (3.133)
and the elements of matrix and are
(rl"'_rS»')(£<'•»))„ if p=i
if p i (3.134)
71
and
( > • ! . » ' - i f q = i (3.135)
i f q ^ i .
During the calculation of the exchange forces, we choose configurations which have a
positive determinant value. .\s we mentioned in section 3.3, a necklace of an electron
must have a closed form and be treated exactly in the same manner both in the harmonic
force and the exchange force calculations.
3.5 Parallel Computation
In the path-integral molecular dynamics of a many-fermion system, most of computing
time ( > 95 % of total time) is devoted to the calculation of the exchange force and energy
of a particle. We have shown in the previous section, that the calculation of the exchange
force can be accelerated by making appropriate approximatins. In this section, we show
that the exchange forces calculation can be implemented on the parallel computer leading
to a significant reduction in computing time.
One often wirtes the program with vectorization algorithm to use a vector-type
computer for a long computational job. The vector-type machine, however, can not be
very useful for this particular simulation, because the inner most routine of our simulation
contains a determinant calculation which can not be vectorized. One other method to
reduce the real computing time (i.e. the wall clock time) for a long simulation is a parallel
computation. Our simlulation algorithm is an ideally sited for the parallel computation,
because we can independently evaluate the exchange forces on P different nodes of an
electron necklace. If we have an M-processor computer and assign P/M jobs to each
process, we can finish the job M times faster.
To see the parallel algorithm clearly, let's consider a physical quantity A
P
(3.136) 1=1
72
where can be the exchange force or the exchange energy of a single bead i of an
electron necklace. Because the caJculation of can be performed independently of the
calculation of any other equation (3.136) can be rewritten as
A = x: + • • •+e A h Im
or M P'
.4= (3.137) 1=1 t=i
where P' = P / M , J [ i ) = i" + (/ — 1)M, and M is a positive integer which is equal to
the number of processors. With equation( 3.137), each single processor only calculates
rather than in a normal serial computer. Figure 3.13 shows
the sequantial order of the parallel computation. First of all, we allocate M different sub-
jobs, to M processors with ail information necessary for the calculation delivered
by a processor which is called the Master. Upon complete reception of the necessary
information each processor sends a signal to the Master and starts starts to calculate
the subjob. When the processor has finished the subjob, it sends send the results to
the Master followed by a job-end signal. When the Master gathers all job-end signals
with the results from the M processors, the Master proceeds to a next sequantial job.
The allocating and the gathering procedures of information takes extra computing time
which is called the communication time. In our simulation, the total amount of commu
nicating data is normally on the order of 1 Mbytes. No data have to be transfered from
one process to another during the calculation. Since the bandwidth of a communication
network is about 10 Mbytes/sec, the communication time is less than 1 % of the com
putational time. In figure (3.14), we illustrate the scalability of computing time by the
number of processors. The computing speed has been increased very sharply up to 30
processors.
To implement the parallel program, we use the Message-Passing Interface (MPI)
[31] library tools which is based on the message-passing model machine. The message-
passing model posits a set of processes with local memory which are able to communicate
with each other processes by sending and receiving message through a communication
73
PO(Master): allocating all information
to M slaves
PO: gathering all results
from the slaves
Figure 3.13: Sequantial job order in a parallel computer.
74
1-0
0.8
0.6
0.4
0.2
0.0 20 40 80 100 120 0 140
Number of Processors
Figure 3.14: Scalability of the parallel calculation. CPU time scale is normalized to 1 for the seriaJ calculation with one processor. All time is measured at the IBM SP2 in the Telecommunication Center at Cornell University.
75
network. MPI is a specific realization of the message-passing model. IBM SP2(in the
Telecommunication Center at Cornell University) and SGI Origin 2000(in the CCIT at
University of Arizona) have been used to run the parallel simulations.
76
CHAPTER 4
APPLICATIONS AND RESULTS
4.1 Electron Pkisma
4.1.1 Model System
We have tested the restricted path integral molecular dynamics on an unpolarized elec
tron plasma composed of NeU = 30 electrons(iV, = 15 with s = spin-up or spin-down).
.A.n uniform positively charged backgroud is imposed to neutralize the system. The sim
ulation cell is a fixed cubic box with edge length £.=13.3A, 19.95A, or 26.6A, which
correspond to electronic densities with 7.5, and 10, respectively. r,=r/ao where
Bohr radius and r = (3/47r7z)'''^ is the electron radius for the electron number density n
and oq is the Bohr radius. We applied periodic boundary conditions in all three dimen-
tions so that the system is constitutied of one basic simulation cell and 26 image cells.
Under these conditions the total potential energy is equal to the sum of the interactions
within the basic cell plus the interactions between the basic cell and all 26 image cells
minus a background term.
With Eq. (2.80) and (3.98), we solve the equation of motion with a leap frog
scheme and an integration time step At = 2.010"^® sec. Most simulations were run for
an average of 30,000 time steps. In some cases for the low and intermediate density
plasmas, we have run simulations up to 50,000 steps for better equilibration.
Because of the large computational cost of the calculation of the exchange effective
potential and forces, the exchange forces are calculated and updated every 10 time
I I
steps. The values for the exchange forces are used subsequently during the 10 time
steps following their calculation. We have compared the average energies obtained one
from the sicip-procedure and the other from the non-skip-procedure, found no significant
statistical difference in their values. We show the energies obtained from both procedures
in Fig. 4.15.
The chosen time step is small enough to resolve the high frequency oscillations
of the harmonic motions related to the potential. In the case of system with large
P, the strong harmonic forces in equation (2.60) may lead to non-ergodic behavior[33].
This problem can be alleviated by rescaling temperature with a necklace of Nos^Hoover
thermostats[49, 50]. This rescaiing would ensure convergence to the right canonical
distribution. We have elected to rescale the temperature of each necklace of P beads
independently of each other via a simple momentum rescaling thermostat[93]. With this
procedure we do not obtain a true canonical distribution, but most thermal averages will
be accurate to orders N~^ [30].
We chose an Ewald parameter t} = for which satisfactory convergence is
obtained with truncation of the real space sum at Z,/2 and truncation of the reciprocal
sum at < 49. As described in section (3.3), every simulation starts with independent
initial configuration obtained from randomly generated bead positions in every electron
necklace. The kinetic energy is calculated with the first two terms of the energy estimator,
Eq. (2.85), because {KEexch) is very small and stable values ( -0.05 ± 0.01 eV/electron
for r, = 5 aX T = 1300K and 1800K) and requires more than twice of computing time
necessary for the exchange force calculation. We also note that the error of the kinetic
energy estimator has been estimated by calculating the standard deviation on the running
cumulative average over the last 15,000 time steps of each simulation. This error is on
the order of 0.03 eV per electron.
78
170
a
§ 150 c 3 •a 0 u 3
1 130
110 4000 2000
Time step 1000 3000
3.0 o
c 2.5 3
- 2.0 3 "O &
O) w 1.5 o c o
1.0 c o o Q.
0.5 0 2000 1000 4000 3000
Time step
Figure 4.15: and potential energy in reduced units as functions of time steps. In both cases, the thick lines and the thin dotted lines refer to skip = 10 and skip = 0, respectively.
79
4.1.2 Results
First of all, we have investigated the convergence of the energies with respect to the
number of beads, P, in an electron necklace. In figure 4.16, we report the kinetic energy
of the high density plasma (r, = 5) at T = 1800 K and the intermediate density plasma
(r, = 7.5) at r = 700 K as a function of the number of beads. The convergence of
the potential energy with respect to P is illustrated in figure 4.17 for r, = 5 at T" =
1800 K and r, = 7.5 at T" = 700 K cases. We note that the energies converge to
some asymtotic value for necklaces containing as few as 200 to 300 beads even for the
electron plasma near metallic density. This observation is particularly significant as the
non-parallelized path-integraJ molecular dynamics algorithm scales with the square of
the number of beads. In additio, to PIMD energies, we have also indicated the 0 K
kinetic and potential energies of electron plasma with same density of reference 16. .A.t
the temperature of 1800 K and 700 K, the high and intermediate density systems are
in the degenerate regime and the electronic energies should be comparable to the T =
0 K values. The kinetic energies are in very good agreement but some discrepencies
e.xist between the potential energies as the restricted PIMD appears to over-estimate
them. In order to further the validation of the restricted PIMD, we have conducted a
series of calculations at several temperatures for the three densities. For the high and
medium density systems we have used 450 and 300 beads, respectively. These number
of beads fall within the region of convergence. Electrons in the low density electron
plasma are discretized over 360, 380, 450, 680, 720, 780 for the temperatures 1100, 900,
450, 575 and 550, 500 and 450, 400 and 350 K. The calculated kinetic energies of figure
4.18 are in excellent agreement with the variational Monte Carlo results of Ceperley [12]
for correlated one-component plasma. We note that the kinetic energy is not varying
significantly over the range of temperature studies as is expected for these plasma at the
border of the degenerate and the semi-degenerate regimes[20]. At low temperature, the
low density system with large numbers of beads takes a very long time to equilibrate
and sampling of phase space is not very efficient. In this case, calculation of reliable
energies require very long simulations. Another difficulty in calculating reliable energies
80
I I f 1 T
200 400 600 800 Number of Beads (P)
Figure 4.16: Kinetic energy of electron plasmas as function of number of beads in the necklace representation of quantum particles. The circles and squares refer to the high density (r, = 5, 7'=1800K) and medium density (r, = 7.5, r=700K).
81
-1.0
200 400 600 Number of Beads (P)
800
Figure 4.17: Potential energy as functions of number of beads. The circles and squares refer to the high density (r, = 5, r=1800K) and medium density (r, = 7.5, r=700K).
82
when large number of bead are used results from the fact, as was noted before, that
the variance of the kinetic energy increases with P. We did not need to use so many
beads for the low density plasma even at low temperature, however, these simulations
illustrate the need to use as few beads as possible within the interval of convergence. In
addition to the T" = 0 K correlated energies, we have indicated the Hartree kinetic energy
(2.21/r^ in Rydberg) with a dotted line. Fig. 4.18 shows that the non-local form of the
density matrix given by equation (2.59) introduces some electron correlation. This is
also apparent in the results for the temperature dependence of the potential energy. The
calculated potential energy falls between the fully correlated results of Ceperley and the
electron-electron interaction contribution to the Hartree-Fock energy (given by -0.916/r,
in Rydberg). We also note that the potential energy increases weakly with temperature
and that extrapolation toward T = 0 K should result in potential energies in better
agreement with the correlated potential energies than uncorrected ones. In the present
model, however, the non-local effective potential introduces electron correlation between
electrons with identical spins only. The present potential energies are over-estimated as
correlations between electrons with opposite spins are not ciccounted for. In (B) of the
table we linearly fit the potential energies by using a -I- bT. In table 4.1, we show the
kinetic and potential energies of different models.
Finaily as a demonstration of the effectiveness of the exchange potential in equa
tion (2.73), the pair correlation foriso-spin and hetero-spin electrons is reported in figure
4.19 in the case of high density plasma at the three temperatures studied. The difference
between iso-spin and hetero-spin radial distributions is striking. In order to satisry Pauli
exclusion principle, the non-local exchange potential keeps the electrons with identical
spin away from each other while electrons with different spins can approach each other
quite closely. The Coulomb repulsive force is the only force keeping electrons with dif
ferent spins from approaching. The non-local exchange potential is quite short range
as it does not appear to affect the electron distribution beyond 5 A. The exclusion is
particularly important in the interval (0, 3 A). The major effect of a rising temperature
is the increase in pair correlation at shorter distance or in other words, the shrinkage of
83
(A). Kinetic energies per electron.
r,(= r / a o ) Current model (eV)
reference 12 (eV)
Haxtree-Fock Approximation(eV)
5.0 1.511 (± 0.045) 1.529 1.203
7.5 0.764 (± 0.022) 0.774 0.534
10.0 0.523 (± 0.076) 0.491 0.301
(B). Potential energies per electron.
r,(= r/ao) Current a(eV)
model b (lO-* eV/K)
reference 12 (eV)
Haxtree-Fock Approx.(eV)
5.0 -3.396 ±0.1367 2.73 -3.619 -2.493
7.5 -2.355 ±0.0306 1.51 -2.517 -1.662
10.0 -1.685 ±0.0474 1.46 -1.944 1.246
Table 4.1: Kinetic (A) and potential (B) energies per electron for various electron densities. Kinetic energies per electron of current model are the average values. Potential energies per electron are linearly fitted by using PE = a + bT.
84
2.0 • I > 1 ' ' 1 r—r—I 1 1 1 1 r
0 500 1000 1500 2000 2500 Temperature (K)
Figure 4.18: Kinetic energy as functions of temperature. The electron plasma with ra=5, r,=7.5 and r5=10 are refered to by circles, squares and triangles, respectively. The horizontal thick dashed lines correspond to the energies of Ceperley[12,13]The thin dotted lines indicate the Hartree-Fock energies.
85
-1.0
-1.5
-2.0
-2.5
-3.0
-3.5
-4.0
» !
I
' •
500 1000 1500 Temperature (K)
2000 2500
Figure 4.19: Potential energy as functions of temperature. The electron plasma with ^3=0, rs=7.5 and rs=10 are refered to by circles, squares and triangles, respectively. The horizontal thick dashed lines correspond to the energies of Ceperley[12,13]. The thin dotted lines indicate the Hartree-Fock energies. For both types of lines, r,=5, r,=7.5 and r,=10 are represented from the top to the bottom.
86
1.0
hetero-spfn c o
W (0
a 0.5 0 1 o
I
I
iso-spin r :
0.0
Radial distance (Angstrom)
Figure 4.20: Iso-spin and hetero-spin electron-electron pair distributions for the high density (r, = 5) electron plasma at T=1300 K (solid lines), T=1800 K (dotted lines) and T=2300 K (dashed lines).
87
the exchange-correlation hole[20].
After showing that the non-local restricted PIMD can simulate with reasonable
accuracy electron plasma near metal density, we apply the method to the simulation of
the solid and liquid phases of an alkali metal from first-principle.
4.2 Soiid/Liquid Transition of Alkali Metal
4.2.1 Model System
We simulate crystalline and liquid potassium metal with method of the restricted path-
integral molecular dynamics. Potassium has been choosen because
(1) it is a prototype free-electron metal which has been studied previously by semi-
empirical pair potential,
(2) there exist experimental data for the pair-correlation function of the liquid state[89],
thermodynamic[3] and vibrational properties[19].
The alkali metal system is composed of 54 potassium ions (K"*") and 54 unpolarized
electrons in a cubic basic cell with edge length L. The number of electrons with spin-
up and spin-down is N-a,-p = 27 and N^onm = 27, respectively. In the solid phase, the
potassium ions and the electrons are arranged on a body centered cubic (bcc) lattice.
We start every simulation with a newly generated random configuration. In contrast
to the electron plasmas, when ion temperature is varied, the dimensions of the basic
cell are adjusted to match the experimental density of K crystal. The density of K
crystal is linearly varying with temperature; D{T) = Dq+ [T — Tojo:, where Dq = 0.827
gjcrr? is the density at the melting point Tq = 337 K and the expansion coefficient a = -
0.2285 x].Q~^g/Kcm^ [76]. Unexpectedly, our simulation have shown that the potassium
model system melts at a temperature below the experimental vaJue of the melting point,
consequently the density of the liquid system reported below conforms to the value of
the density of crystal.
88
Periodic boundary conditions (PBC) are applied in three dimensions to the basic
simulation ceil and the cutoff range of potentiaJ. is chosen to the half size of the basic
cell, i.e. £/2. We use Au = 0.6172 x 10^° eV and pu = 0.1085 A for the Born-Mayer
potential for the ion-ion interation, Eq. (3.101) [72], a core radius Rc = 2.22 A [38] for
the ion-electron pseudopotential, and the Ewald parameter 77 = 5.741/Lo . I.0 = 16 .4..
The physical potassium ion mass is Mi = 71,830 me, leading to an extreme
disparity in electronic and ionic time scales. For practical reasons, we use a ratio of
the ion mass M/ to the electron bead artificial mass, m*. of 39.1 ; 1, i.e. m* = 1
amu. The dynamics of the electrons is still significantly faster than the dynamics of the
ions. We solve the equations of motion with a leap frog scheme and an integration time
step At = 2.8 X 10"'® sec. Most simulations were run for a minimum of 70,000 time
steps (~ 20 psec). In some cases for the calculation of vibrational properties, we have
run simulations upto 130,000 steps. As we described in section 4.1.1 for the simulation
of the electron plasma, the exchange forces are calculated and updated every 10 time
steps. Then we use the updated exchange forces for the next 10 time stef>s to reduce the
computational cost.
We have studied the potassium model system at temperatures in the interval
[lOK, 298K]. The simulation of the electronic degrees of freedom as discrete necklaces
as these low temperatures would necessitate a large number of beads for convergence
with respect to P. From the study of an electron plasma at high density (near the metal
density of potassium), the electron systems conserves a nearly degenerate character up
to a temperature of 2300K. Since temperature does not affect significantly the electron
states at the metal density, we have thermally decoupled the classical ionic degrees
of freedon and the quantum elctronic degrees of freedon. The electron necklaces are
attached to a thermostat set as a temperature of 1300K while the ionic temperature
is adjusted independently with another thermostat. At the electron temperature of
1300K, it is sufficient to employ a reasonably small number of beads for convergence of
the algorithm.
89
The calculation of the electron kinetic energy is done in a similar fashion to that
of the electron plasma. With the kinetic energy estimator, the kinetic energy is given as
a small quantity, difference between two larger quantities, with a variance growing with
P. This estimator, therefore, introduces an error on the calculated values of the average
kinetic energy. We have estimated this error by calculating the standard deviation on
the running cumulative average over the last 30,000 time steps of the simulations. This
error is estimated to be on the order of 0.01 eV per electron. Figure 4.21 shows the
running averages of the electron kinetic energies at T = 273K and T = lOK.
4.2.2 Results and Discussion
In a first stage, we have investigated the convergence of the algorithm with respect to
the number of beads in the electron necklaces, namely P. For this we have calculated
the electron kinetic energy at an ion temperature of 273A' and an electron temperature
of 1300A' for systems with varying values of P. It is important to note again that each
simulation starts from different initial necklace configulations. Figure 4.22 presents the
results of these calsulations. It is clearly seen that the electron kinetic energy converges to
an asymptotic valuse of approximately 1.23 eV/electron. The algorithm appears to have
nearly converged for number of beads exceeding 240. As a trade off between accuracy
and efficiency, we have choosen P=260 for all subsequent simulations.
The total energy of the potassium system as a function of temperature is reported
in figure 4.23. The energy shows two regions separated by an apparent discontinuity of
approximately 0.025 eV/atom. In figure 4.23, we have also drawn as a guide for the
eyes best second order and first order polynomials fits to the low temperature and high
temperature energies, respectively. The slope of the curve increases significantly from
the low to the high temperature region indicative of larger energy fluctuations in the
high temperature systems. In table 4.2, we compare the slope of Fig. 4.23, i .e. AE/AT,
and the gap of the total energy at T = 230K with the experimental values.
90
(A). Specipic heat
SE/ST (Current model) C„ (experiment) (10""* eV/atom K) (10""* eV/atom K)
Solid 1.810 ± 0.0145 2.540^^^
Liquid .3.786 ± 0.0144 3.048(2)
(B). Latent heat of fusion
AE at 230K ~ 0.025 eV/atom
AH (experiment) 0.0245 eV/atom
Table 4.2: Specific heat of the solid and the liquid, and latent heat. The experimental values are at the data book(CRC). (1) and (2) in (.A.) are measure at T = lOOK and 298K.
It shows that the vibrational modes of the model metal are softer than their
experimental counterparts.
We recall that the density of the simulated potassium system varies continuously
as a function of temperature as it is set to the temperature dependent density of the
solid. Therefore the discontinuity is not associated with any discontinuous change in vol
ume of the system but can only result from a structural transformation. This structural
transformation takes place around 210A'. .\s we will see later from structural data, this
is a solid to liquid transformation. The calculated transformation therefore underesti
mates the melting point by nearly 12GR' as potassium melts at SSS/C under atmospheric
pressure. This difference cannot be assigned to the fact that the density of the simulated
system is constrained since such a constraint should have the opposite effect of raising
the melting point. The difference between experimental and simulation melting point
can only result from the computer model that underestimates the strength of the K-K
bond and in particular we believe that it is a consequence in part of the approximation
made to reduce the range of the Coulomb interaction. In that respect it is predominantly
a size effect.
91
To gain further insight into the energetics of the transformation, we have graphed
in figure 4.24, some of the contributions to the energy of the system. The only energy
that is not plotted is the classical kinetic energy of the ions. Since the temperature of
the ions is maintained constant by a thermostat, the ion kinetic energy is a simple linear
function of temperature and cannot account for the discontinuity in the total energy.
.\part from an isolated point at 200A', the potential energies vary reasonably continu
ously with temperature. In contrast, it appears that the electron kinetic energy data
is separable in two groups, namely a low temperature group and a high temperature
group. Since the electrons in the potassium system are nearly degenerate, their kinetic
energy should not be temperature dependent provided the atomic structure remains the
same. Within each group the kinetic energy does not show any systematic variation. We
should remember that the standard deviation on the electron kinetic is approximately
0.01 eV/electron. The difference between the energies of the two groups amounts to
approximately 0.015-0.02 eV/electron and appears to be a significant contribution to
the total energy discontinuity. The raise in kinetic energy as one crosses the disconti
nuity from the low temperature to the high temperature is indicative of a change to an
electronic state of higher localization in the high temperature metal. This observation is
consistent with the expected behavior of electrons in a liquid structure in contrast to a
crystalline solid. As the structure disorders from crystalline to liquid, one anticipates a
narrowing of the electronic band. However, since the short-range local atomic environ
ment does not chanfe drastically between the liquid and the solid above and below the
transformation temperature, the e.xtent of the electronic localization should be small.
We characterize the atomic structure of the simulated system via the ion-ion pair
distribution function. The distributions calculated at several temperatures are drawn in
figure 4.25. The very low-temperature ion-ion pair distributioni function shows a first
nearest neighbor peak at approximately 4.6A and well-defined second nearest shoulder
at 5.3 A. The third nearest neighbor peak occurs near 7.4 .4. These features are char
acteristic of the body centered cubic structure of crystalline potassium. As temperature
increases, the second nearest neighbor shoulder fades away and merges with the first
92
nearest neighbor peak forming a broad asymmetric peak because of the large amplitude
of atomic motion. At the temperature of 76A', the third nearest neighbor peak retains
its identity. At 150^, this peak consists only of a vague shoulder part of a much broader
peak that should encompass higher order nearest neighbors. However, due to the limited
size of our simulation cell, we cannot resolve with much confidence the pair distribution
function beyond one half the length of the edge of the simulation cell. On the same
figure, we have also plotted the ion-ion pair distribution functions at the temperatures
of 248A" and 273K. The distributions at 273A' and 298A' are practically identical. The
maximum of the first nearest neighbor peak shifts toward lower values as temperature
increases. At 273A", this maximum occurs at a distance of approximately 4.3 .4. This
distance is an underestimation of the experimental first nearest neighbor distance of the
ion-ion distribution of potassium [89] but the calculated liquid distribution is in good
qualitative accord with available experimental data.
To supplement the structural information provided by the ion pair distribution
functions, we report in figure 4.26 two-dimentional projections of the trajectories of
the potassium ions at several temperatures. The first two figures,4.26 (a) and 4.26
(b), correspond to the crystalline states. The ionic species vibrate about well defined
equilibrium lattice positions. At the two high temperatures, fig. 4.26 (c) and 4.26
(d), one cannot identify lattice positions anymore. Although one may still identify
some vibrational component to the ionic motion in the form of some localization in the
trajectories, ionic motion is not predominantly oscillatory but also possesses a diffusive
charater.
More quantitative information concerning ionic motion is available from the
analysis of the mean square displacement. Figure 4.27 shows the mean square displace-
ment(MSD) of potassium ions as a function of time and temperature. In terms of the
MSD, diffusive motion is identified by linear variation with time in the limit of large
time[5l]. Vibration motion is characterized by a time independent MSD. At the three
lowest temperatures(10, 76, and 150A), the MSD indicates that ionic motion is vibra
tional. At the highest temperatures of 248, 273 and 298K, the ions exhibit diffusive
93
motions. It is somewhat more diflScuIt with the temperature of 223K. However, because
the density of the system is constrained to conform to that of the solid, it is not surpris
ing that at 22'3K, atoms in this liquid may display essentially vibrational motion. We
also illustrate the vibrational amplitudes of ions as a function of the ion temperature in
figure 4.28. The vibrational amplitude is taken as the maximum amplitude experienced
by an individual ion during its trajectory. Individual amplitudes are then avetaged over
the ions and reported in figure 4.28. In this figure, the vibrational amplitudes of ions are
almost linearly increasing up to T =223K. The slope of the dashed line is 6.33 xlO~^
A/K. In this low temperature region, the increase in amplitude may be associated with
an increase in the number of phonons. A simple dimensional analysis predicts that the
amplitude increases as y/T in classical temperature regime. Because the data of figure
4.28 are so qualitative one cannot predict accurately the actual temperature dependence.
At T = 250K, we can still estimate a vibrational amplitude owing to the partial vibra
tional nature of ionic motion. This amplitude exceed the crystal value by as much as a
factor of 2, showing a structive transformation.
Further information on the ionic motion is obtained from the calculation of the
normalized velocity auto-correlation function(NV.A.F). We also consider the power spec
trum of the NVAF. defined as its Fourier transform. The NVAF's and associated power
spectra have to be analyzed in a qualitative manner because the time over which they
are calculated is not long enough for quantitative characterization. In figures 4.29, the
NVAFs at the two high temperatures of 21ZK and 298/v" show features of the crystalline
state such as oscillations representative of thermally excited phonons in crystal lattices.
The contrast in ionic motion between the liquid and the crystalline is also quite apparent
in the power spectrum and in particular in the low frequency modes. At lOA', the power
spectrum drops to zero at zero frequency. The liquid systems at 273A' and more evi
dently at 298A', exhibit non-zero values of the power spectrum at zero frequency. This
observation is in accord with a diflFusive ionic motion [78], 2044(1992)]. The peaks in the
power spectra of the liquid metal are consequences of the oscillations in the NVAF's and
may thus be regarded as remnants of the phonon structure observed in the crystal state.
94
The fact that the density of liquid is constrained to that of the crystal may accentuate
this effect. .As temperature increases or density decreases, these peaks should disappear
with the decay of the oscillations. It is not possible to extract detailed information from
the fine structure of the power spectra because of the finite time used in their calculation.
However, one may compare qualitatively the calculated power spectrum at lOA' with that
deduced from experimental measurements at 9 A' [19]. The experimental phonon density
of states possesses a major peak near 2.1 x Vibrations in the PEMD model of
the crystal potassium have lower frequencies in the range 0.8-1.3 x suggestive of
weaker bonds. This observation correlates closely with the observation of a calculated
melting temperature underestimating the experimental melting point.
Finally, we consider the change in electronic structure of the metal upon melting.
This change is associated with an increase in electronic kinetic energy of approzimately
0.02 eV/electron. This energy is small and thus one expects only a slight modification
of the electronic structure. Such a variation is obserable in the electron pair distribu
tion function of figure 4.31. The partial pair distribution functions show that the major
difference between the low temperature crystal and the liquid is an increase of the max
imum in the hetero-spin pair correlation between 3A and 4A and an ezpansion of the
exchange-correlation hole as seen in the iso-spin distribution. In a previous study of the
effect of temperature on the electron density in an electron plasma with near that of
the present potassium system, we had shown that increasing temperature shrinks the
exchange-correlation hole[58]. However, the direct effect of temperature in the electronic
dtructure cannot be a factor as it is maintained constant by a thermostat. Here, the
e.xpansion in the parallel-spin electron correlation may thus simply be a result of vol
ume expansion. On another hand upon melting the first nearest neighbor and second
nearest neighbor shells of the crystal structure collapse and the ion coordination num
ber in the liquid increases. The exchange-correlation force between neighboring iso-spin
electrons may then induce further localization. The resulting localization within and at
the border of the ionic core is seen best in the electron-ion radial distribution of figure
4.32. At low temperature the ion-electron pair distribution function shows a significant
95
first maximum at a distance of 2.2A. This distance corresponds to approximately one
half the first nearest neighbor interatomic distance. The ion-electron correlation reaches 9 9
a minimum between 4.3 and 4.5 A followed by a second maximum near 6.5 A. The
ion-electron pair distribution is therefore complementary of the ion-ion distribution. In
other words, high electron-ion correlation is expected where there is low ion-ion correla
tion. The high electron density between ions is indicative of bonding. Considering now
the mid-points between ionic sites as consisting of electronic sites, we can estimate the
electron-electron distance in the potassium bcc structure to be on the order of L/\/2
times the lattice parameter. The electron-electron distance thus calculated amounts to
approximately 3.67 A. This number is in e-xcallent accord with the observed maximum in
the hetero-spin electron-electron pair distribution. With this information, we may con
struct a simplified picture of the electronic structure in the crystal phase. The electron
density is highest between the ions thus leading to bonding and the electronic sites are
occupied alternatively by electrons with differing spin.
An increase in electron localization at the electron sites occures even in the solid
state at the higher temperature of 150A*. This shows that atomic vibrtations have a
significant effect on the nature of the electronic states in crystalline potassium. Vibra
tions tend to localize the electron density.Similar observations by other investigators were
made for the case of sodium clusters[35, 36]. The electron density localizes further with
disordering of the structure at even higher temperature. In the liquid, the electron den
sity increases near 2.2 A. This increase is compensated by a reduction in the electron-ion
pair correlation at longer range as seen by the loss of electron-ion correlation near 6.6 A.
Since the calculated cynamical properties support the retention of vibrational motion in
addition to diffusive motion in the liquid state, it is nuclear at this stage which of two
processes: vibration versus disorder, contributes principally to the localization.
The electron localization at the edge of the ionic cores should leaA to an increase
in hetero-spin correlation between 3 and 4 A which is observed in figure 4.31. Finally, we
note that the larger electron density between nearest neighbor ions is consistent with the
shorter ion-ion bond length in the liquid structure. In order to see the electron necklace
96
with ions, we add figure 4.30. Althought it is hard to tell anything quantitatively, we
still can find that the most beads remains between ions and spreads over several ions.
97
1.24
T = 273K
1.22
O)
.A>v f
T = 10K S 1.20
1.18 40000 60000 50000 70000
Time steps
Figure 4.21: Running averages of electron kinetic energies at T = 273K and T The standard deviations are 0.003 (eV/eiectron) and 0.005 (eV/eiectron) at T and at T = lOK, respectively.
= lOK. = 273K
98
1.40
J -Si >
>. O) h. o c o o
c o o o lU
1.20
1.00 -
0.80
I
100 150 200 250 Number of Beads (P)
300 350
Figure 4.22: Convergence of electron kinetic energy with respect to the number of beads (P) in potassium. Tion = 273 K and Teu = 1300 K. Nion = = 54.
99
50 100 150 200 250 300 350 Temperature (K)
Figure 4.23: Total energy of the potassium model versus temperature. The lines are fits to the data in the low and high temperature regions.
100
E o o > 3-O) w O C o «
o a.
TJ c a •
I •¥
-2.30
-2.35
-2.40
-2.45
-2.50 100 200 300
Temperature (K)
o 1.24
100 200 300 Temperature (K)
100 200 300 Temperature (K)
Figure 4.24: Various contributions to the total energy of the potassium system as functions of temperature.
101
T = 10K T = 76K T = 150K T = 248K T = 273K T = 298K
.O 3.0
4.0 5.0 6.0 Radial distance(Angstroin)
Figure 4.25: Ion pair distribution functions at the different temperatures of 10K(thick solid line), 76K(thick dotted line), 150K(dashed line), 248K(thin solid line), 273K(thin dotted line), and 298K(thick long dashed line).
102
% >
• r
t
. « 4t %
(b)
-V
Figure 4.26: Trajectories of the potassium ions at (a) T=10K, (b) T=76K, (c) T=248K, and (d) T=298K.
103
0.06
T=298K
273K
0.04
Q CO S
0.02
223K .
150K
76K
10K 0.00 10
T(10 '^sec)
Figure 4.27: Mean square displacement (MSD) of potassium ions as a function of time and temperature.
104
i
50 100 150 200 Temperature (K)
250
Figure 4.28: Vibrational amplitude of the potassium ions as function of the ion temperature. The dotted line is a linear fit to the data except for the last point.
105
UL < >
T=298K
T=:273K a 0.04
vw\/\Ar^Ay||| S 0.02
T=10K
0 12 3 Frequwieytio" Hz)
T( IO "sec ) Frequency (10 Hz)
Figure 4.29: Normalized velocity autocorrelation function(NVAF) and associated power spectrum for crystalline potassium(T=10K) and liquid metal(T=273K and T=298K). The insert in the T=10K power spectrum is the experimentally deduced phonon density of state at T=9K of a reference [19].
106
% * • • t
Figure 4.30: A 2D projectioa of the electron necklace (open circle) with potassium ions (large filled circle) at T = 298K and T = lOK. The frame represents the simulation cell.
107
1.2
1.0
antiparallel spin
c o 0.8 3 n
parallel spin
(B Q. o • 0.4 o
I
I
0.2
0.0
Radial distance(Angstrom)
Figure 4.31: Partial electron-electron pair correlation functions. The solid lines and dotted lines refer to the crystal at T=10K and the liquid at T=273K, respectively.
108
1.2
1.1
1.0
73 0.9
0.8 T=10K T=150K T=273K
0.7
0.6
Radial distance (Angstrom)
Figure 4.32: Ion-electron pair distribution functions at severaJ temperatures.
109
CHAPTER 5
CONCLUSIONS AND FUTURE WORK
We have shown that a non-local form of the restricted discretized path integral may be
used to define an effective exchange potential for use in molecular dynamics simulations
of quantum particles obeying Fermi statistics. A quantum particle is represented as a
closed necklace of discrete beads. Exchange is described via non-local cross-linking of
the necklaces. First, we have demonstrated that electron plasmas may be simulated
with a satisfactory degree of accurjicy with this method up to metallic densities. We
have noted that the exchange potential appears to introduce correlation in some effec
tive form. We have applied the first-principles molecular dynamics scheme to the study
of the finite temperature properties of a simple metal. We showed that the model potas
sium metal undergoes a melting transformation upon heating. The transformation is
characterized thoroughly through calculated thermodynamics quantities, structural and
dynamical properties. The model potassium crystal melts at a temperature significantly
below the experimental melting point. The reasons for this discrepency may be found in
the approximations made to speed up the calculations as for instance the use of a short-
range interatomic potential and an empty core pseudopotential. Comparison between
the calculated low-temperature vibrational spectrum and the experimentally measured
phonon density of states indicates that the strength of the metalic bonds is underes
timated in the model. Upon melting the ionic motion changes from pure vibration to
diffusive motion. Above the transition temperature, the ion mean square displacement
increases linearly with temperature an unambiguous sign of diffusion. This change in
atomic motion is also supported by the temperature dependency of the Fourier trans
form of the ion velocity auto-correlation function. The path-integral MD allows us to
110
study the interplay between the atomic and electronic structures. In the crystalline state,
vibrations appear to have an effect on the electron density and result in some electron
localization. Moreover, we find that the electronic structure of the simple metal responds
to the collapse in long range order of the ionic structure by localizing within and at the
edge of the core the ions.
Contrary to many of the current quantum molecular dynamics simulation tech
niques which rely on the independent particle approximation, the path-integral MD is a
many particle technique and includes the important effects of interactions of electrons
with each other and with the ions. Although path-integral MD a very promising tech
nique for the study of materials in which electronic and ionic structures are intimately
correlated, the shear computational cost of the algorithm constitutes a barrier to its
application to large systems. At present, the restricted path-integral MD method is
limited by the computation cost of forces derived from the effective exchange potential.
The computational cost is a quadratic function of the number of beads and a cubic func
tion of the number of isospin electrons. Access to supercomputers can make possible
the simulation of systems with larger numbers of electrons. For larger fermion systems,
one may be able to optimize the calculation by exploitng the short spatial extent of
exchange[47] and dividing the simulation cell into smaller and more tractable subcells.
The quadratic dependency on P due to the non-locality of the exchange potential is a
more serious problem. A local effective exchange potential could lead to a linear de
pendency on the number of beads. We are currently developing an approximate local
form of the exchange potential that is able to model the exchange interactions at a cost
proportional to P only[61].
Ill
CHAPTER 6
APPENDICES
6.1 APPENDIX A : FREE PARTICLE PROPAGATOR
In this Appendix, we will prove equation (2.16)
We assume that every state is defined by a set of plane waves. The free particle propa
gator in the position representation can be written as
PQ(ri,rj;e) = (r,| e-'^|r_,>
= J <ip{ri\p){p\e-'P
= j rfp<»-.|p)<p|rj>e-"'^/^'". (6.139)
We used a completeness of momentum states, 1= /dp\p){p\, in Eq. (6.139). If we use
a plane wave basis (r,|p) = ^^"^pT^-exp (ip • r,/^), Eq. (6.139) becomes
4n- /•<» , sin (or;,/ft)
where Vij = Jr,- — rj|. From an integral table, we have
xsin {tx) e~ '^dx = t (6.141)
Thus if we set a = £/2m and t = we have Eq. (6.138).
112
6.2 APPENDIX B : EXCHANGE KINETIC ENERGY ESTIMATOR
In this section, we will ccilculate the exchange kinetic estimator, {KEexch)r which is given
by Eq. (2.86),
(o down P P f \ I (««(£"")) s=up k=l 1=1 /
/dovm . P P 1/5 \
In order to differentiate a determinant, we use the following matrix aigebra: if a .V x N
square matrix A is function of X, we have
N Af det A = 51 Otj Aij = Oij Aij, (6.143)
.=1 j=i
^detA = (6.144) i=l j=l
where a,j is an element of the matrix A and A,j is a cofactor of the element a,j. Thus
if we differentiate det(£''^''') with respect to /?, we have
= EE I (£"•"),. .4, «=1 J=1
^ JL J _ «(0^2 _ ,^(fc) _ J0^2^ U . = E E Hp ((r!" - .yv - (rl" - }]. ,
= iSi E E ((-•!" - -y - (--l" - ••!")') (£"•"). • 'ly- (6.145)
If we apply Eq. (6.143) to the left-hand side of the last term of Eq. (6.145), we can wirte
the equation with a simpler form;
!<!«(£"•')) = <'•'>), (6.146)
113
where is an iV X iV matrix and its element is given by
exp i f i = t
{^n.= (6.147)
and
i(W) = ri') - r<'V - (rl'l -
From Eq. (6.142) and (6.146), the exchange kinetic estimator can be written as
(dovm , P P f iV p„ \ E ^ E E E ««('f;"') • (S.148,
6.3 APPENDIX C : EXCHANGE FORCE CALCULATION
Here we calculate the exchange force. From equation (3.127),
fW _ i. ^ g(£'(**''^))py «("•")) C6 149^ '3EE<,et(£:(».'))EE ( )p» (6-«)
where
= exp{Co((rM)2-(rl'-l)2 + 2r;,''>T(''>-2rJ"l-rM)} (6.150)
with Co = and is a cofactor of a matrix element Using Eq.
(6.150), then we have
5r- '
-rjr^Si,pSk,^ - r^Si^pSk,,) (6.151)
The first two terms of the left-hand side of Eq. (6.151) with (6.149) become
e(f:),lst,2nd Pm njp-Ak},iH,2nd ^ y" (Ji. <5i. J t,exch ^2^2 ^ P "t,q°k,l/)^pq Opg
114
^ ^tkl^ f (k) (n,k) ^ ^2^2 I ^ ' '' ^l'' '"'"
= 0. (6.152)
In Eq. (6.152), we used the relation det(£'^^''*') = det(£'^'^'^)) and the matrix algebra
Eq. (6.143). From the third and the fifth terms , and the fourth and sixth terms of
equation (6.151), we have
P P a+ ip» S' N'
Jt,exch ft2fi2 2^ 2^ (1&t( ^ f P t.P"k,ii)^pq Opq
Pm ^ &tJP' _-(>'Up(i',u)Mk,u) , /32^2 det(£:(*''')) '' ' (b-15J)
and similarly
P P fl+ /P- 'V' N' Ak)Ath,6th ^ fm ^ _-(/x)x. X,
•'i.excA jg2 2 Z-f 2^ \ 2-f P P t-P"k,u)£'pq "p, a=l "=1 ^ ' p=l 7=1
P™ - - J:
^p=l p=l
£(r('') - rS''>)£'Jr'*^B<f (6.154) M=1 p=l
Form the above equations, the exchange force at the A:th bead of the ith electron becomes
fl'l^ = If S d^i£i) } (6-155)
where the elements of matrix pj and (jJ"'*' are defined as
(rl"l - r|"'){E('-»))„ if p=i
;32ft2 2^ M=i det(f;(''.*))
Pm p
Kk/p-^2ft2 2-r
/t=l det(E(»-'>)
Pm P »lklP-
= •(
, (£"'"')p, if P¥^i-
and
(£;('''%, if q=i
(6.156)
(G1»'")„ = (£<'-*l)„ i f q ^ i
(6.157)
115
REFERENCES
[1] A. Alavi aad D. Frenkel, J. Chem. Phys. 97, 9249 (1992).
[2] A. Alavi, J. Kohanoff, M. Parrinello, and D. Frenkel, Phys. Rev. Lett. 73, 2599
(1994).
[3] C.B Alcock, J. of Physical and Chemical Reference Data, 23, 385 (1994).
[4] B.J. Alder and J.E. Wainwright, J. Chem. Phys. 27, 1208 (1957),
[5] B.J. Alder and J.E. Wainwright, J. Chem. Phys. 31, 456 (1958), and ibid, 33, 1439
(1960).
[6] H.C. .A.ndersen, J. Chem. Phys. 72, 2384 (1980).
[7] J.E. Anderson, J. Chem. Phys. 63, 1499 (1975), ibid 65, 4121 (1976), and ibid 73,
3897 (1980).
[8] P.E. Blochl and M. Parrinello, Phys. Rev. B 45, 9413 (1992).
[9] S.G. Brush, H.L. Sahlin, and E. Teller, J. Chem. Phys. 45, 2102 (1966).
[10] D.J.E. Cailaway and A. Rahman, Phys. Rev. Lett. 49, 613 (1982).
[11] R. Car and M. Parrinello, Phys. Rev. Lett. 55, 2471 (1985).
[12] D.M. Ceperley, Phys. Rev. B 18, 3126 (1978).
[13] D.M. Ceperley and B.J. Alder, Phys. Rev. Lett. 45, 566 (1980).
[14] D.M. Ceperley, J. Stat. Phys. 63, 1237 (1991).
[15] D.M. Ceperley, Phys. Rev. Lett. 69, 331 (1992).
116
[16] D.M. Ceperiey, Phys. Rev. Lett. 73, 2145 (1994).
[17] D.M. Ceperiey, Rev. Mod. Phys. 67, 279 (1995).
[18] D. Chandler and P.G. Wolynes, J. Chem. Phys. 74, 4078 (1981).
[19] R.A. Cowley, A.D.B. Woods, and G Dolling, Phys. Rev. 150, 487 (1966).
[20] R.G. Dendrea, N.W. AshcrofFt, and A.E. Carlsson, Phys.Rev. B 34, 2097 (1986).
[21] P.P. Ewald, Ann. Physik. 64, 253 (1921).
[22] R.P. Feynman, Phys. Rev. 76, 769 (1949), ibid 90, 1116 (1953), ibid 91, 1291
(1953), and ibidQl, 1301 (1953).
[23] R.P. Feynman and A.R. Hibbs, Quantum Mechanics and Path Integrals (McGraw-
Hil, New York, 1965).
[24] R.P. Feynman, Statistical mechanics, (Benjamin, New York, 1972).
[25] L.D. Fosdick and H. F. Jordan, Phys. Rev. 143, 58 (1966), ibid 171, 128 (1968).
[26] W. Freyland, Phys. Rev. B 20, 5104 (1979).
[27] F.G. Fumi and M.P. Tosi, J, Phys. Chem. Solids, 25, 31 (1964).
[28] M.Gell-Mann and K. Bruekner, Phys. Rev. 106, 364 (1957).
[29] D.A. Gibson and E.A. Carter, J. Phys. Chem. 97, 13429 (1993).
[30] M.J. Gillan, in Computer Modelling of Fluids, Polymers and Solids, C.R.A. Catlow
et al. Eds. (Dordrecht, Kluwer Academic, 1990), p.l55.
[31] W. Gropp, E. Eusk, and A. Skjellum, Using MPI: Portable Parallel Programming
xnith Message-Passing Interface, (MIT Press, Cambridge, 1994).
[32] M. Haile and S. Gupta, J. Chem. Phys. 79, (1983).
[33] R.W. Hall and B.J. Berne, J. Chem. Phys. 81, 3641 (1984).
117
[34] R.W. Hall, J. Chem. Phys. 89, 4212 (1988).
[35] R.W. Hall, J. Chetn. Phys. 91, 1926 (1989).
[36] R.W. HaU, J. Chem. Phys. 93, 5628 (1989).
[37] R.W. Hall, J. Chem. Phys. 93, 8211 (1990).
[38] W.A. Harrison, Pseudopotentials in the Theory of Metals (New York : Benjamin)
1966.
[39] M.F. Herman, E.J. Bruskin, and B.J. Berne, j. Chem. Phys. 76, 5150 (1982),
[40] D.W. Hermann, Computer Simulation Method in Theoretical Physics, (Springer-
Verlag, Berlin, 1986).
[41] W.G. Hoover, A.J.C. Ladd, and B. Moran, Phys Rev. Lett. 48, 1818 (1982).
[42] C. Kittel, Elementary Statistical Physics, (John Wiley & Sons, New York, 1967).
[43] C. Kittel, Introduction to Solid State Physics ed. (Wiley, New York, 1992).
[44] H. Kleinert, Path Integral Quantum Mechanics, Statistics, and Polymer Physics
(Singapore : World Scientific, 1990).
[45] A.J.C. Ladd and W.G. Hoover, Phys. Rev. B 28, 1756 (1983).
[46] C.Y. Lee and P.A. Deymier, Solid State Comm. 102, 653 (1997).
[47] X.P. Li, R.W. Numes, and D. Vanderbilt, Phys. Rev. B 47, 10891 (1993).
[48] P. Linse and H.C. Anderson, J. Chem. Phys. 85, 3027 (1986).
[49] G.J. Martyna, M.L. Klein, and M. Tuckerman, J. Chem. Phys. 97, 2635 (1992).
[50] D. Marx and M. Parrinello, Z. fur Physik B 95, 143 (1994).
[51] M.A. McQuarrie, Statistical Mechanics, (Harper and Row, New York, 1976).
[52] N. Metropolis and S. Ulam, J. Am. Stat, Assoc. 44, 335 (1949).
118
[53] A. Muramatsu et al. Int. J. Mod. Phys. C3, 185 (1992).
[54] J.W. Negele and H. Orland, Quantum Many-Particle Systems, (Addison-Wesley,
New York, 1988).
[55] B.R.A. Nijboer and F.W. De Wette, Physica, 23, 309 (1957).
[56] S. Nose, J. Chem. Phys. 81, 511 (1984).
[57] S. Nose, J. Chem. Phys. 81, 511 (1984).
[58] Ki-Dong Oh and P.A. Deymier, Phys. Rev. Lett. 81, 3104 (1998).
[59] Ki-Dong Oh and P.A. Deymier, Phys. Rev. B 58, 7577 (1998).
[60] Ki-Dong Oh and P.A. Deymier, Phys. Rev. B, in press.
[61] Ki-Dong Oh and P.A. Deymier, in preparation.
[62] M. Parrineilo and A. Rahman, Phys. Rev. Lett. 45, 1196 (1980).
[63] M. Parrineilo and A. Rahman, J. Appl. Phys. 52, 7182 (1981).
[64] M. Parrineilo and A. Rahman, J. Chem. Phys. 80, 860 (1984).
[65] M. Parrineilo, Solid State Comm. 102, 107 (1997).
[66] F. Perrot and M.W.C. Dharraa-wardana, Phys. Rev. A 30, 2619 (1984).
[67] E.L. Pollock and D.M. Ceperley, Phys. Rev. B 30, 2555 (1985), ibid 36, 8343
(1987).
[68] W.H. Press et ai. Numerical Recipes in C, (Cambridge Univ. Press, Cambridge,
1988).
[69] G-X Qian, M. Weinert, G.W. Fernando, and J.W. Davenport, Phys. Rev. Lett.
64, 1146 (1990).
[70] A. Rahman, Phys. Rev. 159, 98 (1967).
119
[71] M.J.L. Sangster and M. Dixon, Adv. in Phys. 25, 247-342 (1976).
[72] M.J.L. Sangster and R.M. Atwood, J. Phys. Chem. C 11, 1541 (1978).
[73] G. Senger, M.L. Ristig, C.E. Campbell, and J.W. Clark, Annals of Physics 218,
160 (1992).
[74] A. Selloni, P. Carnevali, R. Car, and M. Parrinello, Phys. Rev. Lett. 59, 823 (1987).
[75] B. Simon, Funtional Integrals and Quantum Mechanics, (Academic Press, New
York, 1979).
[76] C.J. Smithell, Metals Reference Handbook, 7th ed. (Butterworth, London, 1992).
[77] R.M. Stratt, Annu. Rev. Phys. Chem. 41, 175 (1990).
[78] J. Thalhaiier, Phys. Fluids B 4, 2044 (1992).
[79] S. Tanaka, S. Mitake, and S. Ichimaru, Phys. Rev. .A. 32, 1896 (1985).
[80] E. Teller, J. Chem. Phys. 21, 1087 (1953).
[81] M.P. Tosi and F.G. Fumi, J. Phys. Chem. Solids, 25, 45 (1964).
[82] H. Trotter, Proc. Am. Math. Soc, 10, 545 (1959).
[83] M.E. Tuckerman, B.J. Berne, and A. Rossi, J. Chem. Phys. 94, 1465 (1991).
[84] M.E. Tuckerman, B.J. Berne, and G.J. Martyna, J. Chem. Phys. 97, 1990 (1992).
[85] M.E. Tuckerman, B.J. Berne, G.J. Martyna, and M.K. Klein, J. Chem. Phys. 99,
2796 (1993).
[86] L. Verlet, Phys. Rev. bf 159, 98 (1967).
[87] S. Ulajn, .4 Collection of Mathematical Problems, (Intersdence, New York, 1960).
120
[88] W.W. Warren, Jr., in The Metallic and Nonmetallic States of Matter, P.P. Edwards
and C.N.R. Rao, editors, (Taylor and Francis, London, 1985).
[89] Y. Waseda, in The Structure of Non-crystalline Materials (McGraw Hill, New
York) 1980.
[90] S.R. White, Phys. Rev. B41, 9031 (1990).
[91] S.R. White, Phys. Rev. Lett- 69, 2863 (1992).
[92] W.W Wood and F.R. Parker, J. Chem. Phys. 27, 720 (1957).
[93] L.V. Woodcock, Chem. Phys. Lett. 10, 257 (1971).
[94] L.V. Woodcock and Singer, Trans. Fraday Soc. 67, 12 (1971).
[95] S. Zhang, J, Carlson, and J.E. Gubernatis, Phys. Rev. B 55, 7464 (1997).
IMAGE EVALUATION TEST TARGET (QA-3)
150mm
IM/IGE. Inc 1653 East Main Street Rochester, NY 14609 USA Phone: 716/482-0300 Fax; 716/288-5989